Recsperts - Recommender Systems Experts

Note: This transcript has been generated automatically using OpenAI's whisper and may contain inaccuracies or errors. We recommend listening to the audio for a better understanding of the content. Please feel free to reach out if you spot any corrections that need to be made. Thank you for your understanding.

We as a community should move towards having recommendation systems as a real support structure, not just as a predictive tool.
The psychology in front recommender systems are systems that integrate cognitive, emotional and personality-aware modeling into the design of the algorithms.
Humans are not just data points, but they are humans with goals, with memories, with emotions and with cognitive abilities and cognitive limits.
Psychology can help us model these kind of complex aspects of human behavior and human preferences in a much better way.
What does it mean if some algorithm is more accurate than the other on a specific metric?
Is this really more helpful for the user?
Does this really reflect the user's preferences in much better ways?
User experiments also from a psychological perspective.
It's not only measuring conversion rate, it's about asking the questions about the user's attitude, about their opinions, about their motivation, about their perception, about their lived experiences when they work with an algorithm.
If the user interface is not great and the algorithm can deliver the best results, the user will already be biased towards being more negative about the results because the user interface is maybe overly complicated or does not show why certain recommendations are compiled for them.
Hello and welcome to this new episode of Recsperts, Recommender Systems Experts.
For today's episode we are going to talk about psychology-informed recommender systems, which are combining psychological findings with the recommendation process.
We will also discuss more practically how cognition constructs can be used effectively to enhance prediction of user re-listening behavior and hence generate better recommendations.
This whole discussion might then also bring us to some of the very first recommender systems whose publication lies back almost 50 years.
And for this episode I'm very happy that another very great expert has joined me for this discussion and I have invited Professor Dr. Elisabeth Lex from Graz University of Technology.
Hello Elisabeth, welcome to the show.
Hello, it's very cool to be here and I'm excited to talk about this topic that is so close to my heart in this podcast.
I hope that many in the community find it very interesting.
I hope so as well.
It nicely adds to all what we are doing and throughout the episode we will also learn like this is not totally a new topic but something that we could also argue everything has started with but more about that to come soon.
So as always I will start with a very brief introduction of my guest and then hand over to you Elisabeth.
So Elisabeth is a professor at Graz University of Technology.
Her postdoctoral thesis was on modeling and predicting user behavior in web-based systems.
She received her PhD in computer science also from Graz University of Technology in Austria in 2011.
And she has published more than 100 scientific publications in many different venues such as the web conference, of course ReXS, UMAP, ECIR, Izmir and many more as well as on different journals.
And she has also been serving the community as a chair and also as a committee member for different positions across the ReXS conference and also along other conferences.
And as we already said like the thing or the research that lies most close to her heart is actually about psychology and from recommender systems.
But not only but I guess Elisabeth you're the better person to talk about your research, what brought you there, what inspired you to delve into recommender systems and those things.
So please tell our listeners about yourself.
Wow what a great introduction already.
Thank you.
So we really made your homework.
So I was always interested in supporting users and supporting people with information and with the best data they can have.
So this was always my motivation.
And this also brought me to the field of recommender systems research quite early on.
I started in my PhD also to work on information retrieval systems but as we know both communities have their synergies and their differences.
One issue that I saw with recommender systems research also with IR research is that many approaches she uses just as data points.
So we collect data.
Of course data is great and we love it.
But in the end we sometimes forget about that humans are not just data points but they are humans with goals, with memories, with emotions and with cognitive abilities and cognitive limits.
And there I think the integration between data driven approaches and psychological insights is really beneficial to help us address this kind of gap between humans not being just pure data points but also people with different interests and abilities.
This was kind of the motivation for all of my work.
And in the end so I of course want to make algorithms smarter but my overarching objective has always been to make people smarter in the sense that they have really access to all the information that they need in order to make good decisions.
And you mentioned so nicely about my CV.
So I mean I am currently an associate professor but I'm also the incoming professor for HCI and inclusive technologies at my university starting very soon.
And there I'm very excited to bring this topic of recommendation systems of user modeling and personalization also to a community where there are people really with different abilities and different limits and needs.
I think this is a lot of untapped potential there where specifically the integration of psychology and data driven algorithms can be very beneficial and lead to new services and ideas because users are not just clicking machines they're not just data points but they are people who forget, who get distracted, who regret.
And psychology can help us model these kind of complex aspects of human behavior and human preferences in a much better way.
This human nature might be something that from a computational or computer science point of view we might perceive as noise but then it's not just noise it's psychology and it's human beings as you said like which have different attention or different memory, different mood and I guess this is also some of the subfields within that where you talk about effect aware recommender systems where mood plays a big role.
So something that recommender scientists, engineers, however we want to call them have always somewhat assumed being there or captured with different things but not really taking the more holistic picture of that's really inspire or work by psychology.
Yeah absolutely and I think even so in the beginning of the community and your thesis really so I think there was more awareness on this.
It means that the first recommender system that you mentioned in the beginning that was invented or that was proposed and published I think in 1979.
Actually this was a paper or a system that was proposed by a computer scientist but it was published in a cognitive science journal.
So Alain Rich proposed the Grundy system which was a recommendation system I think for TV based programs if I remember correctly and it employed the idea of stereotypes as a way to categorize people according to their preferences.
So you had like stereotypical preferences and then you could get recommendations based on those stereotypes.
And while we in many discussions we think that stereotypes can be a really bad thing in the end if we discuss this from a psychological perspective it's just a heuristic that humans apply in order to make faster decisions.
So it's a psychological concept that reduces complexity exactly and that can help us create our personalized recommendations.
I think so in the early years in our community there was a lot of awareness that there are disciplines that have a rich body of insight about human behavior and human preferences and how they are formed.
But then of course machine learning came into play and all these very very great algorithms and the abilities to collect data and to find patterns in data and I think this is where we got lost a little bit on where these communities kind of like moved into different directions.
But we see this now in the RecSys community a lot that we have increasingly more researchers trying to bridge again the disciplines to make more human centered algorithms a reality.
And I think this is very good because it means every digital service that we now interact with has some form of recommendation algorithms behind it.
I think they are really everywhere and in the end they are always humans and they are humans with different needs like I said before with different preferences.
So I think we as a community should move towards having recommendation systems as a real support structure not just as a predictive tool and for that psychology is really one domain where we can tap into to get good models and good theories about user behavior.
And maybe then also a point where both areas of fields recommender systems and psychology actually intersect because psychology I mean I'm not a psychology expert you are definitely better to tell me more about it as it's your research but I assume like one field of psychology is also trying to understand possibly also to model and explain like how people make decisions and what drives their decision making and one of the definitions for recommender systems it's seen as a system that should enable humans making decisions in the abundance of information and there you can already see like this is not separate it's intersecting and when it's intersecting why cannot tools or approaches from the one domain enrich the goal that the other is trying to solve.
So it's definitely like very logical but I found it also interesting that you said this increasing abundance of data the availability of especially behavioral data has somewhat moved us away.
Why do you think this is the case or why has that happened?
I'm not completely sure why this happened but I think when your home is in different scientific disciplines let's put it like this you realize that each scientific discipline has their own norms, has their own language, has their own methodologies and bringing reaching those is not an easy task right.
It's hard on its own to become a machine learning expert for instance right so you need to invest a lot of time a lot of energy a lot of you need to acquire a lot of knowledge in order to do that and the same holds true if you want to become an expert in psychology right so you need to study a lot of theories you do need to learn a whole set of methodological tools in order to become this and having both at the same time from is quite challenging I would say also for me it was quite challenging so when I started working in this field I also had to start reading psychological papers right on the one hand you need to find out what papers to read how they are structured and how to understand them because simple words can have very very different meanings already so for instance the word model in machine learning it's clear right what yeah you have an algorithm to create a model whatever but the word model in psychology has a slightly different meaning so we as computer scientists we always try to make abstractions we are very happy to make abstractions like to like lose some of the necessary information but in other domains this is seen very critically so you want to build a model that really captures all of it and everything of it so I think this is why this is one of the reasons why these communities drifted apart and also in terms of like science landscape so there are not a lot of incentives to do to work in an interdisciplinary fashion it's harder to get research funding it's harder to get yeah papers published because you need to have them people's reviewing those papers who can appreciate both sides of the story right so I think those were like like barriers to this and at the same time yeah I think the success of machine learning of deep learning really yeah sparked a lot of large developments in the community that are focusing mostly on optimizing algorithms which is great yeah of course but it's easy when you have just data that you think okay everything is in the data I just need to have better algorithms to find those patterns to disregard all the theory that is behind human behavior for instance this actually remembers me of one sentence that Robin Burke mentioned and I guess it's where he said it's appealing to solve an easy problem and the availability of data gives us somewhat the at least appearance that it's easier to solve the recommendation problem yeah which can mean many things but let's keep it as such so like oh there's a lot of interaction data of our behavioral traces and now solving for the recommendation problem should be way easier but in the end I'm neglecting stuff as I'm just switching focus from one thing to the other and jumping on what seems as a more easier problem to solve yeah I think this is a perfect explanation and I think this is what we what we can observe yeah so no harm done with this kind of approach but I think we can do more and we can do better and we need to do better since recommendation algorithms are present in so many systems now and you have so different users interacting with the system you have small children you have people with different disabilities you have people from different genders and all of them so reflecting all of this in data is then on the other hand again a hard task because typically you reflect certain demographic in a data or a certain piece of the world in the data not everybody is is there so then if you would have done a personalization algorithm that which should work for a community that is not as well reflected in the data then your system will have a hard time targeting those people yeah I found also when preparing for this episode practical applications of your research that you and your fellows did on predicting music re-listening behavior since this is quite an acknowledged behavior of users and how to model that better by leveraging psychological constructs hey folks quick pause this is Marcel I started for experts almost four and a half years ago since then I've released more than 30 episodes with guests from industry and academia reaching listeners in over 54 countries across all platforms and over the lifetime of the show we have crossed 50,000 place with listeners mainly in the US Germany India the UK and the Netherlands and we are close to 1,000 subscribers on Spotify alone which is great I run this podcast completely voluntarily there is no monetization just a lot of time spent researching reaching out preparing interviewing and producing all to give the recommender systems community a platform and to give you new ideas and perspectives and here's the honest ask despite nearly 1,000 Spotify subscribers there are only 37 ratings on Apple podcasts it's even worse three ratings for about 350 subscribers I'm not asking for money I'm not asking you to join a newsletter I'm not even asking for five stars I'm just asking for 30 seconds to rate the show honestly whether that's three four or five stars and maybe leave a comment on what you like or what could be done better that small action makes a huge difference for discoverability and for keeping this project going thank you to everyone who has already rated the show shared feedback or especially joined me as a guest and now back to the episode when was that first touch point that you had with recommender systems in your career and where was that point you said hey I want to perform research there and dive deeper know and understand more and where was that actually or when yeah that two events that come to my mind one is was in the context of my PhD studies where somebody at my department had a project on learning analytics so the idea was to set up a system where you observe how people interact with learning resources and then to measure the behavior so collect a lot of behavioral data in order to support the people with and recommendations on documents or maybe on people to ask so if they if a conflict could be observed or if somebody was stuck or something along these lines and there I found the application for recommendation algorithms very appealing yeah because it happens frequently so you'll help you have to study something you learn you explore a topic but you do not know exactly so what research should I look at next or who could I ask so this was one of the touch points where I really found the application very very interesting and very relevant and I also like started to work on using recommendation techniques in order to improve learning processes already then and also specifically then afterwards in a postdoc the second event that comes to mind this is kind of a controversial and maybe I should not say this on the RECSPERTSs but this was like this was earlier so it was all early in my PhD when I was at a conference I think it was a hypertext conference or something along these lines and there I saw presentations about recommender systems papers and they were all like algorithms reporting small improvements on movie lens on the movie lens data set and all those presentations were kind of similar so all offline studies on the same data set more or less and reporting like minor accuracy related improvements and there I thought to myself okay so this cannot be it you know so I thought to myself so I mean if this is the type of contribution is this field already solved this is what came to my mind when I saw these presentations and then already then it sparked the idea in my mind that yeah it's it's probably not enough to just improve algorithms on a data set by some some fraction of performance metric right and there I figured either the field is already solved or there is more to do on the user side and then I decided to work on the user set more because I mean also the recommendation algorithms that I interacted with early on were not giving me the best results in many aspects and they still don't do it sometimes so one of my research fields in the direction of RecSys is also to improve our recommendations for people with non mainstream preferences so when you listen to music for instance that is not very popular or you want to watch very niche movies because I personally do that and sometimes I still receive recommendations that are too popular for my taste yeah and this kind of sparked my interest in improving recommendations for people who might not have that stereotypical behavior that many systems assume you have so basically their niche user receiving blockbuster recommendations yeah if you listen to like very specific types of metal music and then you receive nickelback or something along something like that yeah you know what I mean so I had I had exposure to RecSys research quite early on I think but I was conflicted on on my view on the systems but as a scientist I think this is a good thing because it gives you the opportunity to to do research and to study those systems in order to to study exactly those those questions that you think systems are lacking cool then let's just dive into that very topic and lay out the landscape a bit in 2021 you together with other researchers published a survey paper called psychology informed recommender systems summarizing the state of art in research but also I guess this fueled a lot the tutorial that you together with Marco Schiedler gave at RecSys 2022 which was a tutorial on psychology informed recommender systems can you help our listeners with the landscape of psychology informed recommender systems so like which categories to exist how do we like split the different fields there yes psychology informed recommender systems are systems that integrate cognitive emotional and personality aware modeling into the design of the algorithms at least this is how how we categorize this so we looked really at a whole body of literature in different fields and came up with this categorization cognition informed recommender systems use models from human cognition as a basis to design and improve their algorithms and human cognition is is a concept from psychology where the idea is to develop a comprehensive model of how humans perceive the world learn interact and communicate with each other so those are really like very large comprehensive theories but we found cognitive concepts in different stages in the recommendation process while compiling this literature research a huge category of psychology informed recommender systems is naturally emotion aware recommender systems because you mean so specifically when you're thinking about entertainment domain right when we interact with music with movies they they trigger emotion in ourselves but at the same time we also select them based on some emotional values right so for instance during Christmas or so you might select festive festive festive music in order to have this emotion that you maybe connotate with with with Christmas this is a third category we identified personality of where recommender systems which you which you already mentioned as your first touch point on the topic very interesting and this is also a huge part in the community where you take various personality approaches models and tailor recommendations whether based on whether somebody is for instance an extrovert or an introvert or these aspects the idea is really to like represent various aspects of human factors take these models in a way that you design recommendation algorithms in so that they are more human centric and human aware in our survey we not only discussed design of algorithms but also the evaluation and I think their psychology plays a huge role because in order to like set up an evaluation a user centric evaluation also psychological constructs can help you a lot there are user centric evaluation frameworks where you can then test the impact of various psychological constructs on the user satisfaction for instance on the way people appreciate the recommendations this is what we all described in there I think one topic that comes to mind when thinking about psychology and from recommender systems is also the topic of biases cognitive biases this is also captured quite extensively in our paper on the one hand cognitive biases are something that does not cause by recommendation algorithms this is something that happens in our mind those are these mental shortcuts such as heuristics that we use in order to make decisions but at the same time recommendations algorithms can really play into those biases if you think for instance about confirmation bias which is a very well-known cognitive bias and if you design a recommendation algorithm that always just amplifies what the user already knows or what the user already agrees with then you would like your fuel confirmation bias with the system these biases are really hard to study they're hard to identify but at the same time they are very critically when you want to design recommendation techniques that are beneficial for for users and for the society and that upholds certain values with the techniques all right from the perspective of designing recommender systems and joining forces between like you also I guess throughout the survey call it data driven approaches to our recommender system so something that I guess most listeners would refer to this classical collaborative filtering content-based or hybrid methods thereof to join them with psychological models and you already said like there are these three categories so the cognition inspired ones the emotion or effect awareness you also call them in there in the survey and the personality aware shall we just get started with the cognition inspired and talk about them and what constitutes cognition inspired recommender systems I guess you already said there are certain models in cognition psychology that we could use but also you mentioned the term model is used differently across those domains of psychology and computer science so what does cognition inspired mean specifically and could you possibly give an example absolutely as a cognition inspired recommendation systems take models from cognitive psychology or cognitive science to design algorithms cognitive psychology itself is a huge field within psychology and they are the ideas to investigate human mental processes such as decision-making we discussed this already but also memory or attention and those are very important processes when users interact with a recommendation artifact or with a recommendation system right because those are decision-making tools and our attention of course is also influenced by what we see and what we are offered to buy by a recommendation technique or by literally any type of digital service so there are many models of human cognition I would say what's really nice is that cognitive scientists typically take a very computational perspective when they design their models so this is in itself cognitive science already an interdisciplinary field where you have classic AI research in there neuroscientists in there psychologists in there because it's about understanding the human mind and of course this is also something that is very interdisciplinary but this computational metaphor is very helpful of course because there are so called cognitive architectures that are designed to simulate parts of human cognition or maybe also the overall human cognition through computational models so there are psychologists or there have been psychologists that created basically computational models with mathematical formalisms in there to describe processes like human memory processes or like how people remember certain aspects based on what they have learned so far and this is the nice right the model in cognitive science that for instance we have used quite a lot is based on a cognitive architecture which is called ECTR and it's a model of human memory and there you have different memory modules such as the declarative memory which stores like our the episodes that you learn this is more like the long-term memory that we have also in our mind then there is another memory module which is called procedural memory that stores like if then else constructs more or less yeah so if you have learned something like riding a bike then the procedure on how what steps you need to follow in order to ride the bike are stored there so if you're like sit on the saddle then you need to take the pedals and stuff like this yeah so it's like really this these procedures on accomplishing certain tasks there's also another module that has the short-term memories where all the memories are stored for only a few minutes so it's very very intuitive in the sense what we know about maybe from biology classes about our memory and the nice thing is cognitive scientists have then researched these memory processes in empirical studies in simulation studies and came up with really mathematical descriptions of how humans acquire information how they store the information and how they retrieve the information if we're thinking about it creating a recommendation technique is also similar you have some data you store it somewhere you have then so those are maybe our user item interactions and then you retrieve an item based on what the user has liked previously so you retrieve it from memory so this is this is kind of a similarity in the in the process i would say in our work on cognitive inspired recommender systems we then used such memory models in order to model what a user has consumed in the past and what items a user will consume likely in the future based on human memory theory there are also other cognitive models that are very helpful models of attention because users have also shifts in their interest shifts in their preferences shifts in their attention when they interact with items in a system their cognitive models of of attention can help you to see whether an item fits to the current attentional focus of a user for instance like when when when having a smaller or higher attention span or whatever exactly yeah more deserving something exactly lightweight just for for some for some smooth entertainment or whether you want to like yeah see some more complex tv show or or a movie yeah or in the context of learning for instance if a user is interacting with a learning system and the user is currently very focused to learn a concept in very detailed you do not want to like disturb the users with recommendations that are un-similar or that are not not close enough to the attentional focus of the user right while you might have another user who is in an exploratory learning phase you might want to recommend items that are a little bit more diverse right and in order to identify in what state the user is you can use these psychological concepts and psychological models because those shifts in attention they happen quite fast if you would do this in a pure data driven way you might not be that fast enough in updating updating your model i see i see and and one of their good or useful representatives also for your work was this uh ac no was it the first time i was reading what that acronym stands for i was a bit wow okay that could be a quite debatable translation like the adaptive control of thought ration exactly yeah exactly so it was uh yeah okay fits well into the domain of recommender systems so i guess we we will we will learn more about it in the papers where you made use of that actual framework or cognitive construct you said there are different mechanisms or subfields like you said there's attention and there's memory are there also models who do not only specialize on one of those but try to combine these different aspects of cognition there are but they have not really been tested at least to my knowledge in the context of recommender systems of course from the psychological perspective those theories really try to be as comprehensive as possible so they try to be really unified theories of the mind but i think if you would bring this to reality like to a real system where you want to implement a recommendation technique currently we're trying to focus a little bit more on the single constructs in order to understand them better and in order to also combine them with classic approaches right because in the end so while i'm an absolute advocate for psychology at the same time i'm also a realist and of course the performance of machine learning based approaches is also undepatable right so my approach is rather not to like substitute one for the other but just to complement both approaches right and i have not found a good way yet to do this like on a very comprehensive way but rather on a focusing on one construct one aspect of psychological model and then combine it with a high performance machine learning approach in order to like still preserve the capability of the model while being able to fuse in more human centric perspectives yeah makes perfect sense because actually that very example of these two things memory and and attention remembers me okay but it's actually both things that are for example driving my capability of retrieving actual facts from my memory possibly so i guess also during the tutorial you brought up that example of meeting new people at a conference and then you get introduced you hear their name and then of course it's here it's a matter of attention how much attention you you pay to that very situation to the name but it's also like how well your your memory is capable of retrieving that so it's like it's individual aspects but in the end like what drives the outcome of how well you are able to recall or in the first instance store that bit of information is driven by attention and by memory but nevertheless it makes makes perfect sense to first start out like small if you can even say so and understand the individual components before merging them further and then merging is already your starting point because you do not like use just the cognitive construct to model recommendations but you already combine it with with the recommendation model and there is already like something yeah in that combination process that needs to be understood exactly yeah exactly more exactly but at the same time i think there's a lot of potential in this direction that you have now mentioned and one of the research directions that i see is very fruitful is to identify first what kind of cognitive constructs are at play at all right because we make these assumptions maybe we find some evidence for them in data that we observe or in user studies but at the same time we still are yeah not in the situation where we can definitely say okay this construct is at play while this construct is at play while that is this that construct is at play and i mean i have i'm part of a large research project that runs in austria where all the major austrian universities are working towards hybrid combinations of artificial intelligence and i of course work on hybrid combinations of artificial intelligence for recommendation systems naturally with great colleagues such as makoshe lardit mayanach and there the idea is to use symbolic ai symbolic artificial intelligence in order to identify evidence for such constructs right and symbolic artificial intelligence for the ones who are not familiar with this term so the idea is to have or to extract explicit rule-based representations of knowledge and those rule-based representation of knowledge could be an indicator of what psychological construct is at play currently right because we find a rule that says okay so if the person has a high attention focus then they like whatever type of resource and there we also our approach is also deeply grounded in psychological theory but this project has just started so we're just scratching the surface at the moment with it but yeah i see a lot of potential in shedding light on what contracts are even so looking forward for more to come there sure what about personality aware as a field within psychology informed rex's what is this actually about i mean personality is one of the basic human characteristics and psychologists have looked into personality aspects for for many many years personality traits are typically quite stable over the years right and you see this when you have small children you might see evidence already for some personality traits in those children early on of course certain aspects might develop but yeah this is this is quite quite stable which is nice for a recommendation task of course because we love stable preferences and stuff we love stable stable characteristics right yeah what's also interesting about personalities typically personality it means it does not change depending on the context you're in yeah you have your trait and this is more or less fixed or on the stimulus that you're exposed to and there have been many studies that showed that certain user preferences user characteristics have a strong correlation with personality also in the context of a recommendation scenarios there are certain personality traits that correlate with users preferences for diversity for popularity for serendipitous content i mean this is kind of obvious when you think about it when you're a very extrovert person you might enjoy a lot of very different experiences and of course then the recommendation algorithm can nicely account for this while if you are very introverted of then you might have a very particular specific taste when you have a neurotic personality trait some people might have then you might be even more particular of what you want to enjoy from a recommendation algorithm this was kind of the motivation so why to use this in a recommendation scenario and you pointed it out so nicely even there these traits you do not infer them from behavioral data i mean you can but it's not great but there are certain service and certain standardized tests that you can do in order to infer what kind of personality traits are at play in your life so you can take those ask users to fill them out and then you have this data and you can use this in a cold start scenario nicely before the user has made any kind of interactions in your system so what you are alluding to there is for example this ocean profile for instance yeah for instance this big five yeah yeah the big five so i guess it's openness conscientiousness extraversion ex-liability neuroticism exactly yeah exactly and yeah you can take like as you said standardized tests that then put you somewhat in the range what if i and it also remembers me back to of that or um yeah cambridge analytical scandal where this has been used or claimed to be used for and then there was this research like trying to categorize you in that five dimensions based on your likes on facebook and then one could argue yet that with so many likes i can better categorize you as your closest friends someone else like who knows you quite closely for example your family members or certain family members on the one hand side and i guess this is what might come first to many people now listening to this like branch of the conversation is yeah should we really do this this is very critical this is very sensitive data and it tells me a lot about people but on the other side we do have interest to cover how people are or what they are and how they make decisions so like the incentive of using this kind of data to provide better recommendations what i'm thinking about those people who take those tests and maybe it was a central provider and would it be possible to say i use this central repository where i've been taking my test and share that data of course under my control with a platform where i'm recipient of recommendations like for example in music streaming or in video streaming or in news or in a social network whatever how do you think about that like would this be even possible and how critical or less critical i mean you're posing that as a possible direction so and yeah it could be for the sake of just thinking and reasoning about it but why not or why use it yeah sure i mean those are very very good points but they are not only true for psychology informed algorithms but they are true for the whole field of psychology are not only for personality based but also for the whole field of psychology informed recommender systems because you work with very sensitive data very personal data naturally i think i always have the problem that you say okay because it's difficult from an ethics perspective let's just not do it right because in the end as specifically from a research perspective because companies are doing it like you hinted so nicely so we should also research the issues and make clear also the capability of those tests how accurate is this does it even improve the recommendation performance because if it does not then we don't have to use it but it does my intuition is that if we use these type of very personal information we need to ensure that the systems that we develop stick to certain values right that's would be my perspective on this and we as a community have to agree on on those values right and i think in the community we already have some values looking at the papers and you mentioned robin burke so we have fairness as a value we have diversity as a value and these kind of things also i believe i strongly believe in the freedom of of the individual so the scenario that you described so that you have an individual who wants to upload their data to some platform and say okay i want to have this personalization for my context based on my personality traits because i believe this is really helpful to me then why not allow this right there are lots of regulatory procedures in the european union we are here in europe that safeguard this information right um there is the gdpr there is the artificial intelligence act where you have then a lot of regulations in order to ensure that this data is not used in a malicious malicious way and or manipulative form and way so this is my approach to this in the end we also if we just look at user item interactions just behavioral data we already know a lot about people yeah maybe it's not called personality trait from the behavior of people alone we can infer a lot of very sensitive information sensitive preferences right so it's not that different in the sense that we're using user specific data not on a psychological level but on a behavioral level yeah that remembers me somewhat of a discussion that i had where we talked about gender as a sensitive attribute and how just possible it is to use ratings for items to infer the gender here which intuitively makes sense if you assume that there are certain differences in gender and that gender differences might drive differences in interest here and i mean my like my interactions are just a reflection of my interest or like taking that with all the noise in between so you mentioned values on the one side and proper regulation on the other side but this being in place or taking that as provided one should or could be able or enabled to upload that information because and i see that point you were you were mentioning we are already doing this just that we currently use a proxy like i'm not telling explicitly hey this is my ocean profile yeah but on the other hand here are the 100 songs that i listen to or here are the people i follow on linkedin whatever and this is kind of more like okay it's it's more like an inefficient way of eliciting my profile possibly yeah and i think this might be a controversial perspective i'm curious to hear what also others are thinking about this but at the same time i mean we need to just be much better in terms of preserving privacy in ensuring that our data is secure and private i think there more work needs to be done yeah and also to ensure that data is used in a meaningful way so for certain applications you might not need that kind of hyper personalization to specific trade but there are other use there might be some use cases where this is really helpful for you and you should then be able to do this yeah i think this is not something that can be answered on a very very generic level i mean i have a good colleague who is a security researcher he always tells me now don't upload data anywhere that's the only way to ensure the data is secure but i think yeah this cannot be it because then there is no way to achieve personalization and if you're thinking about like very sensitive domains where like in the health sector where you want to get a recommendation that is very very accurate but at the same time you're providing very sensitive data i mean so not have a recommendation there would be not helpful for the user likely but yeah data needs to be secure so it's always a kind of conflict in the trade-off yeah that example sounds to me a bit like don't ride a bicycle if you never want to get involved into a bike accident exactly yeah exactly then then you run over by a car yeah yeah yeah yeah okay so so you said these personality traits they are quite stable something that we like because we don't additionally have to account for like the dynamics which can constitute less complex systems but we also touched on this a bit like what if i'm having a good or a bad day and on a bad day i don't want to see somebody i don't want to interact with people want to get what i already know so even though i might be a very open person in general so the last category yeah how is effect driving this and how is this interplaying with our psychology in front rex's exactly yeah so that what what you now described is not personality but it's mood no yeah and effect and mood of course they play an immense role in our lives at least in my life i guess in everybody's life because we are human beings and we of course are triggered by certain stimuli we're triggered for certain maybe bad days and experiences so there we can distinguish between mood and emotion or at least ecologies to this this distinction so mood is typically something that you experience for a little bit of a longer time so you are in a bad mood you typically this does not last for just five minutes but it can be a little bit longer but yeah it has a lower intensity in the sense that it's maybe not as dramatic for you as having a very strong emotional response to something that happens to you yeah so you're very angry because something bad happens and this task lasts for a shorter period of time yeah but like personality they are very important human characteristics that shape our experiences and they unfortunately or fortunately depending on the decision they shape our decision making abilities and capabilities and also in our preferences that and maybe some people know this behavior when you're in a bad mood or that you then tend to overspend in online shopping or something like that so or when you're very happy then you choose very happy songs or when you're very sad you then choose a more relaxed or even sad song and here a recommendation algorithm can help you like shape these emotions that you feel or kind of influence the emotions or the mood that you feel so we know this from the literature on music recommendations or music listening that when people are in a certain mood or they want to get into a certain mood frequently they use music in order to achieve this so when they want to be in a relaxed mood they listen to relaxed music typically sometimes people also listen to like heavy metal music in order to be relaxed i know those people and while those emotions or affect characteristics are different personality from personality traits there have been also studies that showed that there is some kind of correlation in the sense that people with different personality traits also perceive emotions sometimes a little bit differently this can happen so when you have a narcissistic personality for instance and you can experience certain emotions in a stronger way emotion is then something that you can nicely use in a recommendation algorithm when you detect the mood of somebody then to personalize what the person's versus listening is listening to yeah how would you effectively do so integrate mood or emotion into rex's so is it before using your music streaming platform like look into your smartphone and have an image of your face and we try to detect whether you are in a good or in a bad mood and then like reshape your your recommendations based on that or asking you are you in a good or a bad mood today so so how actually to capture the signal and how to integrate it into the actual algorithm so this is something that i would think is this is quite quite difficult yeah yeah it's quite challenging to do this it means uh inferring this from a facial picture i think it's very very noisy i would not yeah recommend doing this um it was a very invasive in terms of privacy if we come back to this fabric of privacy but there are tools where you can explicitly elicit the mood that you're currently in also in the context of learning there are some certain tools because you want some music can really positively influence learning learning processes and you can then say okay i'm in a good mood i'm in that mood and then you can tailor the music recommendations for instance to this um or it's by selecting and we know some music streaming platforms i'm not gonna name names here but we know some music streaming platforms who have this already in their interface so they they group musical items according to moods and then you can select this okay energetic relaxing and then you can of course observe the user behavior uses behavior on a scale to see okay maybe at that time of day people are more prone towards i don't know having a relaxed mood like on the weekend and then you can push this higher on the interface and there are certain platforms to do this but it's not easy to to elicit this in a responsible and reliable way because people and some i mean i think it's easier with emotions yeah to to clarify for yourself in what kind of emotions that you are at the moment but it's not so easy with mood yeah to see what kind of mood are we really in yeah because it's a very yeah it's a it's not a it's a continuous spectrum i would say right sometimes huh so it's it's one thing that the platform offers you to elicit that but the other like to correctly or properly identifying the mood you are carrying yourself or the emotion that you are feeling because i guess this is also something where some people might be better and some people might be less good at i always remember a situation i had when i was asked about what emotion i had or was feeling and i came up with something that as a person told me that's not an emotion and i was like yeah yeah yeah yeah yeah yeah it's also not so easy to come up with this this standardized test standardized tests that are meaningful to many people because i mean all of these attributes are very culture sensitive also right there are certain cultures who have more emotions who experience like who seem to be like italian you know like all over the place northern german yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah and since both emotion and mood are very are more short-term compared to personality um or also cognitive states that we're using in cognition-informed techniques they change of course incorporating these dynamic changes is of course challenging yeah so an emotion this can be can be over very quickly how is this then actually also related to context-aware recommender systems so like how much is this intersecting or different from context in the way we usually think about it like as time or location but maybe it's mood or emotion also context absolutely absolutely it's heavily intersecting and this can be very beneficial in a context-aware recommendation scenario to have the current mood of a person especially when you have scenarios where you recommend items that impact people's effective states right like movies or something like that or generally anything from the entertainment domain because you you can impact the person's mood or change in a person's mood which can be a tricky tricky thing because if you would detect that somebody's in a in a rather bad mood and you would then recommend uplifting music maybe this would be a natural thing where you think okay this is something that people want but at the same time and we also know this from studies sometimes people want to be want to stay in the mood because they need to process some some emotional stuff and they want to do this and they want to continue doing this and like not having something that they feel is trying to influence their emotions they just want to feel an emotion even though that emotion might be depressing but they want to go through or feel that depressing status and yeah and then then you're back reinforcing that for too long then you might have an issue but yeah it could also be like you're driving high user dissatisfaction if your user feels whatever recommender is trying to convince you or turn your emotion into something different exactly exactly and and you know these effective states also have a strong link to memory some in in the way our memory works is kind of like restore our memory units and our mind assigns them to cues like context cues so location and time are very important context cues this is why when you visit a place that you not have been for a long time and you had made some memories there when you go to the place again this memory will come to your mind yeah you have to try this out this really works every time and at the same time emotions are linked to memories right and since emotions can be linked to items when you listen to a certain piece of music and you have a strong emotion you felt a strong emotion when you have listened to the piece of music a while ago and you listen to it again then this emotion can be triggered oh yeah those are really really interesting correlations and connotations between all of those definitely i have a good example there in mind so there are certain few songs in my lives that are heavily linked to very significant events in my life and they can make me feel very sad or like also very happy if i hear or listen to that music it's not only the memory but it's also the emotion that was linked to the event that was forming that piece of memory yeah yeah exactly yeah and that's by the way also something that you can model with the actr model so you can add this associative activation it's called you can add emotion there as a way as an additional trigger to retrieve a memory unit from from the mind and we also use this in the recommendation context and we saw evidence that it worked but at the same time it's really hard to get data about emotions yeah so you need to do this in a more real user study then so it's not so easy to do this in an offline offline study context because we like the data yeah yeah all right you gave me two cues and now i have to make a decision which cue i take as a seat for the next chapter of our discussion so one cue is actually evaluation and human centered design but you also again gave me the cue about the actr framework and i would like to go with the letter and then come back to matters of evaluation and the parts about human centered design so let's get more practical and talk about two papers that you together with your fellows have published one is actually from rex's 2021 and it carries the title predicting music re-listening behavior using the actr framework and it was published at rex's 2021 where you adopted this framework for explaining music re-listening and then also i like to use it can you give us an overview about the paper and the research question and your your findings there sure i'm very happy to in music many people tend to re-listen to the same items over and over again i do that as well depending on the context depending on the time and this is one of the few contexts where you really appreciate this right so you would not go to a book recommendation scenario and buy the same book every time buy the same book yeah so you would not do that you would not appreciate this but in music this is really something that happens quite a lot and we know this also not only from observing the data that we have about music listening behavior but we notice also from music psychology and in music psychology there have been many papers that linked music preferences music consumption behavior to human memory principles right what we also know from music psychology is that and we know this in general so preferences are not fixed so we might assume the preferences are not fixed but preference develop over time so the more you interact with certain items the more this influences your preferences especially in the context of music so the more frequently you listen to something the more positive your attitude will become at some point also functions with food by the way this also works with food yeah so this effect is called familiarity effect and is well studied in music psychology in our work we also observed this repeated behavior repeated listening behavior to certain items and then we thought okay so we we want to look at this familiarity effect and see if we can model it with a psychological model in order to predict items that the user will listen to next right so this is a sequence-based prediction task so we want to predict the next item in a listening session or the next few items in act are i mentioned this briefly so there is one human memory module which is called the declarative memory module and this stores the factual knowledge that we acquire and it also describes the activation of memory units this activation means so if if a piece of information has a high activation it's more important for us in our current context right and this activation decays over time because not everything can be important for everything in our life so we only have a few items in mind that we can easily retrieve that have this high activation so high activation also means so we can remember it easily and this activation depends on two parameters the one parameter is how frequently you have interacted with something in the past and how recently you have interacted with something in the context of music it means so how frequently have you listened to this track maybe already and how recently and if you have listened to this recently and frequently then it has a higher activation but this follows a time dependent decay function so this is a power law function so it decays quite quickly yeah and what was very interesting to us when we looked at users interactions with music so we had users the tracks that they listen to and the timestamp when did they listen to this and then we sorted this and looked at the distribution the temporal distribution of these listening events and we found that when people have listened to a piece of music frequently and recently they tend to re-listen to the same piece of music again quite yeah following this time dependent function so quite quite often so exactly this this idea that is formalized in act are we could observe this in our data that was very interesting to us and what was also very interesting to us is this was specifically true for people who had a non-typical behavior non-typical i mean in the sense that you have preferences that are not in the mainstream of the community so you do not like what everybody likes but you have a very diverse set of items that you like so it's consistent between niche and those blockbuster users exactly exactly yeah so this function fitted the best for those consumers of niche items yeah and what we then did is we implemented this mathematical formalism which is basically just an equation in order to predict what items then a user will listen to next and this worked really well specifically for the consumers of niche items so they seem to have a strong tendency to like to reconstume the same items over and over again i mean everybody has this but they have it the most telling on a bit differently that there is a difference between how to say the diversity of taste and how much people are on whichever popularity end of the of the distribution meaning that the case they are niche users so interacting with content that is more of the long tail distribution of of song interactions does not necessarily mean that they were having a more or less diverse taste in terms of they do have the model works still quite well predicting their re-listening behavior which for me can also be read as they are not more or less diverse because like if they were more diverse then they would re-listen less songs could someone say that or is it am i making a mistake there yeah i mean we still need to look into exact so i'm i'm in my research i'm i'm careful with making statements where i don't have definitive experiments that show me that this is actually true but what we see or what what i think is that they are more disturbed by social influence than the ones who are at the head of the distribution right in terms of popularity distribution so they seem to have a more individual taste individual taste profiles than people in the in the head what i also would like to study now i mean this is something that comes then what i would like to do in this new professorship that i'm taking there have been works that showed or or we know this basically also from psychology that people who are neurodivergent right autistic people many of those people have a strong preference for safety and for stability so i would assume that a recommendation algorithm being able to like capture these preferences of stability of free consuming similar items over and over again would also be more helpful and would result in higher user satisfaction for those users than for allistic users because for an allistic user it could be really tedious to have the same stuff over and over again you would have you would like to have some sort of social influence at some point so you could combine it with collaborative filtering for instance but for autistic individuals the algorithm could perform better from my point of view it's up to study okay so more stability with the proper modeling through anticipating these factors of recency and frequency throughout the actr framework beyond that so this recency and frequency aspects they form the base level component of this activation yes but there are other components and if i read your results correctly then i would say they haven't added too much on top of the performance but nevertheless they also played a role like the spreading component the partial matching component and the valuation component could you tell us a bit more about them and what is the reasoning behind those components sure yeah the spreading component basically models the co-occurrence between items because of course items are not singular in the sense so when you have listened to a specific track of music then another one might come quite natural to you right so there are co-occurring items in a listening session of users that we can observe help them yeah this could be due to playlists this could be due to like tracks on albums of course and this is why i think this plays a role because some users chant really to listen more on the album level while others more on the individual song level the spreading component really played a little bit of a role in our analysis then there's the partial matching component which denotes similarities between items based on some features and depending on what you have in your data what kind of features those could be this component plays a stronger role than the others but i think while we couldn't observe observe strong strong influence i think this is a matter of the data that we have used in our experiment so i would assume that when we have a richer presentation of item features we would have stronger influence of that component as well so it's also a way of incorporating then content-based signals which then could already constitute for a hybrid recommender that leverage at these exactly okay exactly yeah then there is the valuation component and there we have a link again to the topic of effect-aware recommender algorithms because valuation should model effect yeah so emotional features of course i would also assume this has a strong impact but it's a matter of having the data in order to really make use of that component in our model so we did have some emotion features for our tracks but those were users and tags and they were not very diverse so the very simple tags and there was not a lot of value in the type of data that we had at our disposal i would say and i think this is all contributing to why recency and frequency had the most impact on our model because those were the clearest data signals that we had at hand something that i also found interesting because you compared many different combinations of these different components within the act our framework but also with random recommendations also with our modeling just the transition probability and you basically took two points of view into how you evaluated the accuracy so actually one was like how accurately are we predicting the just next item in the sequence and then the other one was how well do we predict the items that are left in the whole sequence so like which are somewhat coming in the in the whole future not just the immediate next and while this was less good and just predicting the very next item it was very good comparatively at predicting the future items correctly right yeah it was yeah it was and we chose this kind of evaluation design also because if you're thinking about it from a practical value so the next item might not be the best one but the the next after and there's a lot of skip events i mean so this also from a data perspective so we could not test for this but if you would have a real world scenario then the recommendation algorithm could be really good and the way then a user might just click away this next item but the item afterwards yeah could then work really well again our users are not clicking machines not data points so they could have all sorts of reason why they would like to have good recommendations in the whole sequence in a whole listening session not only just one item yeah and i mean if you if you think about the use case of presenting the users with a list of next songs coming up then the user does not only pay attention to the very first maybe the most but sees the list and can somewhat judge like the list is more relevant than maybe just the next item or the first item in the overall list yes i guess in this very first paper you already mentioned in the outlook what if we combine this research with traditional approaches towards recommender systems and this somewhat constituted the work from 2023 with the title integrating the act our framework with collaborative filtering for explainable sequential music recommendation where you also have taken a look beyond accuracy metrics and like also into explainability um can you tell us a bit more about that paper and what you did there and what your findings were sure i mean one disadvantage and i hinted at this really of just using an algorithm such as act r is you get the same recommendations over and over again right and this might not be helpful for all of the users while it might be great for some so we wanted to combine it with a more classic recommendation approach to get social influence because this is what you can can get with collaborative filtering or any kind of other algorithm yeah to have more diversity basically in the sense that you're not just seeing what you already know right so we did this we met several combinations of of our algorithms with algorithms that are common in the community like classics so collaborative filtering bpr is one of the approaches that we evaluated a variational autoencoder just to have this combination of psychological model and classic machine learning model i mean in the end so it's nice to just hit the next item with high accuracy of course this is what we want to publish but is this really helpful for the user is that is a different question this is why in many of our papers we look at to beyond the accuracy metrics novelty is one of the metrics that we evaluated because the idea of combining a classic algorithm with a psychology model was exactly to avoid that there is no lot no novelty for the user so it was of course and they also need to measure how novel are the recommendations that the user then in the end sees right same with diversity we assess diversity with respect to the musical genres that we then recommended diversity and novelty are similar in the sense so you want to have just a broader perspective and give the user also the opportunity to explore the catalog more and then of course popularity bias because this is an issue that yeah many recommendation algorithms have they tend to overly recommend the popular items due to the rich get richer phenomenon we know this and in the end we've also wanted to see whether our algorithm also does this and whether these combinations of algorithms do that because this is something that you want to avoid especially when when we like are so proud to have an algorithm that works so well for consumers of niche items we do not want to like introduce another source of popularity bias with them because this would be would be exactly the opposite of what we were looking looking for right and this is why we chose these kind of metrics in order to get a broader perspective on yeah the evaluation the setup was similar so we had a session based recommendation scenarios we wanted to complete the session yeah in this earlier paper that we discussed we wanted to see so can we fill up the session with meaningful recommendations for the user and we compared of course our algorithm with with other baselines in order to see how well do we score on all of these metrics all right and you actually also came up then using these components with better interpretability yes and i found it quite intuitive how you translated those act are naming so like the base level and the spreading component into something that is more tangible especially if you think about music because then you were relating it current obsession yes that's basically for the base level current wipes for the spreading or evergreens yeah for the stuff that you tend to listen to yeah also the collaborative component and then you could basically highlight why has this made its way in there and how to combine these two signals and weight them properly but also how to wait the individual components within the activation and you found it doesn't really help to learn additional weights on top of it right exactly yeah so at least in our setting in our data we could not show that you really really need to do this in order to achieve good results which was quite kind of interesting i think we could also yeah study this in a different scenario different data set to see whether we have similar results there because it was kind of surprising to me as well i have to say yeah okay it's also cool work maybe maybe to close this topic of an actual application i mean you did that within what we do most of the time so within an offline study on collected historical data based on the last fm data set are you aware of any applications of psychology aware RecSys in really an industrial setting i mean it's of course fine that you possibly can't tell us any names but how far have these approaches already made their way into industrial recommender systems or if not like what would be your recommendation for practitioners where to start i mean what was very interesting not gonna name names in talking with different people from the music domain so they they frequently tell me oh that's the reason why so we make we observe this in our data but we never knew that there's a name to it right yeah that i found very interesting good but also very very interesting yeah yes there was an industrial application and this was a paper that was presented at last year's rex's conference and even nominated for best paper candidate and there basically they combined transformers with act are in order to account for this repeated behavior of music consumption in order to like recommend items in the listening session okay so this is a real world application on a completely different data set that we have used that also showed the benefits of this approach i think this is very very cool in the sense that somebody else confirmed basically our intuition that this algorithm works well in order to complete listening sessions so also worth reading and as always will be added to the show notes of this episode a little bit i guess there's one topic that we haven't covered so far i want to give you also the chance to talk about human centered design that you mentioned in the beginning is somewhat important to this research on psychology and from recommender systems and you also mentioned it in the survey paper recommender systems and human decision making and especially like in that fourth chapter is a user centric commander systems evaluation so what are important things to pay attention in that regard yeah absolutely i mean in many of our paper and also in our papers we perform offline evaluation and i think this is good in the sense that we have the ability to compare across data sets and across performance metrics but at the same time we only get a limited perspective because what does it mean if some algorithm is more accurate than the other on a specific metric is this really more helpful for the user does this really reflect the user's preferences in much better ways because this is exactly what we're interested in right and this is what we describe in our survey as being user-centric recommender systems evaluation so to design evaluation procedures frameworks in a way to really capture the needs of users and really capture the responses of users this is of course not a new topic the whole hci community is doing this for decades and also the recommender systems community is doing this for decades but i think it's important to stress this even more that we increasingly need tools frameworks benchmarks don't also study participants in order to be able to accomplish this so accuracy as a metric is not good enough i would say also the beyond accuracy metrics are better maybe or in combination but it's also not everything yeah does not give us a good intuition about the user satisfaction unfortunately as academics it's hard it's hard to have a real bad system where you can then run a b tests where you can observe users where you can ask them a lot of questions so that's kind of always a balance between those aspects but in our paper we outline why it's important because we describe like the user experiments also from a psychological perspective it's not only measuring conversion rate it's about asking the questions about the user's attitude about the opinions about their motivation about their perception about the lived experiences when they work with an algorithm there are certain biases that are important to be aware when you have a user study when you have a user-centric evaluation such as a user might get recommendations that are very different from what they have known beforehand and they can experience dissonance cognitive dissonance right this would be then again an effective response it can trigger a strong emotion so i don't want to see this this is completely different than what i have had in mind and if you do not account for that or if you do not ask the user of why they experience this dissonance you might just get a negative signal and you don't know how to interpret this negative signal is it the quality of the algorithm is it the content that is displayed think about the news recommendation scenario is it just because there's a different opinion that is expressed to mine no so we want to also understand negative signals in a much much better way and this can only be achieved by doing the evaluations in a very user-centric way the way these interfaces are designed to play a strong role on how users perceive the quality of a recommendation algorithm if the user interface is not great and the algorithm can deliver the best results the user will already be biased towards being more negative about the results because the user interface is maybe overly complicated or it does not show why certain recommendations are compiled for them and so forth so there are many many things that need to be taken into account or where you put the most important result because of yeah primacy and recency effects and so forth so a lot of bias is helping also on the user interface perspective that need to be accounted for when you design a system and when you want to ask the users about whether they like it or not and just one metric does not tell you a lot about why the users like it or not yeah yeah it's hard to compress everything into that yeah yeah give you a hint or direction but then you need definitely more cues absolutely yeah we touched a lot of different aspects on psychology informed rex's so these three different subfields the evaluation i guess there's many more things to investigate and too many to cover and like one and a half hour interview but as i already initially mentioned there's a recording of a great tutorial that you gave in 2022 at rex's but you have also been involved with other workshops in the past so for example the rex so good workshop exactly as part of last year's rex's and this year's rex's and this year is okay so it's many many great contributions from the community to our workshop and there the idea is really to design algorithms and systems with the sustainable development goals in mind so topics such as fairness diversity of course very welcome but also sustainability and i think we as a community we need to think about how much energy and water do our algorithms consume and how sustainable are the practices that we employ when we design new algorithms for different stakeholders and communities yeah i remember there were already some great works especially from the research group in at the university of zeden yeah in terms of like the our environmental footprint of training different algorithms so it's nice to hear that there's more to come and that this workshop that has kick started last year will also again take place this year and it's not only the only one when i also look at what we compile together here i also see that there will be a tutorial at sickir and a workshop at umap what are they going to be about exactly so the tutorial at cigar will be about these cognitive perspectives on information retrieval and recommender systems so a huge part of this is devoted to the topic of cognitive biases i just briefly hinted at those but those are decision biases that can be used for the good and for the bad but i think the most important thing is to be aware of them in order to act responsibly on cognitive biases and i will also give a good overview of cognitive architectures so the act are architecture that we discussed today is just one of them but there are several that are very interesting and that could be employed in the recommender systems community from my point of view much more so in this tutorial we will give you really a comprehensive overview and it's also in a great location it's in italy again so you can come for the great food it's in padua so i hope to see many of you there in july so it's perfect summer weather now then at the umap conference in new york city we will have a workshop which is called hyper yeah yeah it's good for germans in austria it's a nod to our to some musical artists in germany that's an hybrid models for for personalization and there the idea is to combine knowledge-based approaches with machine learning based approaches to have more explainable more fair recommendation and user modeling algorithms we already have selected the papers that will be presented i can share a hint that we will have a fantastic program um so i hope to see many of you at the workshop there as well and i think this is in at least in our community the first workshop that really tries to bring together symbolic and sub-symbolic approaches for the task for recommendation scenarios so i'm looking forward to this first edition so we have sickir umap rex's all well lined up for further diving into this and other related topics together with you and many more skilled researchers in those domains all right which other challenges would you foresee or which would you pose as as a major challenge to solve in in these settings for the future i mean in the machine learning and the ai community we now talk a lot about agents right yeah so you have then basically optimization so you don't have to do so much anymore the question is what will this mean for the field of recommender systems i think this is an open challenge but at the same time it can also be a great opportunity to perceive recommendation systems not only as decision making tools but as a support structure to accomplish more than just what do i need to buy next right and i think this is where where our community will yeah will evolve at least i hope that our community will evolve this into this direction to perceive themselves really as support and there it's even more critical to be human centric to be aware of why you do humans act the way they act why are they interested in certain aspects and what do they really need what is their intent this goes beyond observing user item interactions from my perspective right i think this is one of the the huge topics and the other one is of course llm because this is just everywhere i'm not convinced this will i mean i could also see that llms will go away at some point but maybe not for the for the near future what does this mean for our community on the one hand of course it's great to have conversational interfaces so they can really provide us with conversational interfaces they can help us extract data from natural language text they can maybe help us learn more from large data sets by prompting but at the same time i see an increasing trend to using llms for evaluation like to to bypass the hard task of doing a user study you know what i mean and there is the appeal of solving easy problem yeah and there i'm very very critical and i really hope that our community stays critical because those tools they can be great help us but they are very biased they only represent a certain part of the world they are in transparent they're in the hands of large providers in many cases and they are not humans so we cannot substitute user studies and studies with real humans by llms yeah yeah we need to be very very about the self-reinforcing loop that we might enter with that yes yes all right okay so seems like still a lot of work to be done and interesting questions so we are surely not going to run out of further topics to discuss here at Recsperts but for the moment i want to thank you very much for taking the time and the effort to talk about this in so much detail and so insightful and yeah also providing a new perspective and as you already mentioned they were practitioners reaching out to you that are working on industrial systems that saw the signals who weren't aware of the psychological research and constructs behind that and like making or creating awareness of that this is very very helpful for the community so that every researcher every practitioner gets a broader and more holistic picture and takes it into account in their own work and with that then also make the whole field better more effective and serve ultimately what it's supposed to serve and that's humans yeah thank you for inviting me to this podcast i'm very excited to have had the opportunity to speak about these topics um i think it's important to have venues like this to discuss topics on a yeah to have more time to go and to delve into topics so i'm very happy that i had the opportunity to speak about this thank you and could you think about another person that you would like to listen to in this podcast and who would that be yeah probably did you have bart on already oh not yet yeah because bart has something that is fantastic i did not speak about it today but i just because i just found out about it yesterday but it's not only bart it's a whole community of people so do you know this pop rocks yes and i'm about to talk with joe constant ah yeah of my future episodes okay uh and i guess he will be able to go into great detail at least i hope okay okay yeah because that looks really fantastic to me bart bart kneimberg yeah yeah yeah yeah and he would be from my point of view also a great speaker sounds great yeah yeah yeah and he has a lot to say about evaluation sounds cool no great i always like to get like new pointers also maybe existing pointers reinforced so yeah thanks for the recommendation sure yeah elizabeth it was nice talking to you have a wonderful day and see you at rex's probably see you at rex's bye thank you so much for listening to this episode of Recsperts recommender systems experts the podcast that brings you the experts in recommender systems if you enjoyed this podcast please subscribe to it on your favorite podcast player and please share it with anybody you think might benefit from it if you have questions or recommendation for an interesting expert you want to have in my show or any other suggestions drop me a message on twitter or send me an email thank you again for listening and sharing and make sure not to miss the next episode because people who listen to this also listen to the next episode goodbye so