Designing Recommender Systems for Digital Humanities
In this episode of Data Skeptic, we explore the fascinating intersection of recommender systems and digital humanities with guest Florian Atzenhofer-Baumgartner, a PhD student at Graz University of Technology. Florian is working on Monasterium.net, Europe's largest online collection of historical charters, containing millions of medieval and early modern documents from across the continent. The conversation delves into why traditional recommender systems fall short in the digital humanities space, where users range from expert historians and genealogists to art historians and linguists, each with unique research needs and information-seeking behaviors. Florian explains the technical challenges of building a recommender system for cultural heritage materials, including dealing with sparse user-item interaction matrices, the cold start problem, and the need for multi-modal similarity approaches that can handle text, images, metadata, and historical context. The platform leverages various embedding techniques and gives users control over weighting different modalities—whether they're searching based on text similarity, visual imagery, or diplomatic features like issuers and receivers. A key insight from Florian's research is the importance of balancing serendipity with utility, collection representation to prevent bias, and system explainability while maintaining effectiveness. The discussion also touches on unique evaluation challenges in non-commercial recommendation contexts, including Florian's "research funnel" framework that considers discovery, interaction, integration, and impact stages. Looking ahead, Florian envisions recommendation systems becoming standard tools for exploration across digital archives and cultural heritage repositories throughout Europe, potentially transforming how researchers discover and engage with historical materials. The new version of Monasterium.net, set to launch with enhanced semantic search and recommendation features, represents an important step toward making cultural heritage more accessible and discoverable for everyone.
Guest
Florian Atzenhofer-Baumgartner: I wear many hats in DiDip (didip.eu), e.g., of data analytics, machine learning, DevOps, for historical document analysis systems. I also coordinate and help with research infrastructure efforts in DHInfra (dhinfra.at), including a specialized GPU cluster for humanities computing. I hold a BEd in German and English, with my thesis on learner corpora; a MA in Digital Humanities, with my thesis on text similarity - both from the University of Graz. I am also a PhD candidate at the Institute of Interactive Systems and Data Science (Technical University of Graz), where I do research on recommender systems and ranking in digital humanities and cultural heritage contexts.
Transcript
Kyle Polich : Welcome to Data Skeptic a podcast exploring the methods use cases and consequences of recommender systems
Kyle Polich : Welcome to another installment of Data skeptic recommender Systems Today we're getting into recommenders in the digital humanities that's the interdisciplinary field that combines traditional disciplines like history literature philosophy art cultural studies and so on with modern digital tools methods and computational approaches
Kyle Polich : Our guest today Florian works in this space particularly on a platform called Monasterium.net M O N A S T R I U M dot net if you want to check it out It is the largest online collection of historical charters
Kyle Polich : And has millions of medieval and modern historical documents from all across Europe With a deluge of data like that you're gonna need a good recommender system to find the things that you're interested in We talk about how they're building that in today's interview
Florian Atzenhofer : I'm Florian Atzenhofer and I'm a PhD student at the Gratz University of Technology And in addition to that I'm a developer and IT project manager at the University of Graz in Austria
Kyle Polich : Can you share a few details on what you're studying in your PhD
Florian Atzenhofer : In my PhD I'm focusing on recommended systems for the digital humanities and I'm focusing on three specific aspects of recommended systems in the digital humanities with a specific focus on digital archives and that is multi-stakeholder recommended systems and problems associated with this then the information seeking behavior of specific groups of users
Florian Atzenhofer : And third I'm focusing on the specificity of cultural heritage items and the repercussions on recommended systems of that
Kyle Polich : Could we zoom in on what the humanities need that's let's say unique and you couldn't just use a vanilla off the shelf recommendation
Kyle Polich : approach
Florian Atzenhofer : Um well first of all I must say that there is a huge lack of recommended systems in what you can deem a digital humanities application because naturally uh the digital humanities have been concerned a lot with preservation of cultural heritage data Um and they are only slowly gliding into more innovative ways of how to present the data to users so other platforms and users
Florian Atzenhofer : This means that there have not been so many different applications of recommended systems in digital humanities and for cultural heritage Especially for digital archives and we are disregarding here the digital libraries aspect which includes uh for instance scholarly items such as uh research articles When we focus on primary sources so more historical material we're completely in a new field
Florian Atzenhofer : And is not yet tested whether off the shelf models work well or not And I am arguing already before I'm beginning my work in my thorough work uh that there are several hindrances to just using an off the shelf model You have to deal with a more complex data model So the way of how you structure the data the way how you serialize the data and also the routes to that data are wholly different than for instance in more commercial applications
Florian Atzenhofer : In addition to that you have different user needs You're not dealing with a as I said a commercial application You're not dealing with a typical user-item relation but you're dealing with people and users that are much more invested in the application as you
Florian Atzenhofer : would usually be So the stakes are maybe not that high as in other domains but they are definitely complicated and we are only slowly grasping how to uh formalize them and how we integrate these needs into the system
Kyle Polich : When you work in the digital humanities who is your user
Florian Atzenhofer : First of all I would say that the digital humanities cater to researchers in the first place and and generally to the public domain So they definitely target any user that would also be let's say quote uh user of any glam institution that includes galleries libraries archives and museums
Florian Atzenhofer : So it could be any other citizen that would be a valid target user
Kyle Polich : In your domain you have these experts that have I presume a lot of deep knowledge not necessarily so much surf well they have surface level knowledge but they don't need to query for it and maybe you have only a few people working in areas I guess I'm just speculating as to why user item might not be perfect Could you expand on what the challenges
Kyle Polich : are
Florian Atzenhofer : We do have a a specific relation uh that is driven by the fact that we have selected a handful of experts that is uh very knowledgeable about the specific field and there is a discrepancy between those and the more novice users We need to exploit that relation a bit uh for instance leverage the knowledge of the experts
Florian Atzenhofer : For instance for cold start problems in recommended systems That is the problem that you have new users um and you have a lot of items and many of the of them and the very large majority of them have not been interacted with This means that you have a very sparse user item interaction matrix It is even in other domains where you have more interaction between them
Florian Atzenhofer : It is quite hard to predict what what item a user might like and the more sparse it gets the harder it gets This means that um we need to leverage specific methodologies for instance uh how you deal with cold start problems or how you could lead novice users and do that in a grounded way
Florian Atzenhofer : This means that we will consult with the experts and we will also consult a potential novice users in how they would like to be guided how they would like to receive recommendations
Kyle Polich : Very interesting because yeah it's a journey I could see a 1st year researcher asking very different questions than a 4th year researcher and maybe the system could become aware somehow
Florian Atzenhofer : Yes exactly And also if I may add here we're talking about charters here Uh that's where I'm coming from That's where my project is working in or working with And charters are relatively encompassing for medieval times and modern history And we have hundreds of thousands no we have millions of them spread over Europe Yes there's a a huge bias towards Europe And one of the major problems uh considering this is that
Florian Atzenhofer : We are not talking just about historians or auxiliary historians who could be interested in this source because the source is so important to understand past for societies
Florian Atzenhofer : But this is also used by genealogists by art historians because we're dealing with illuminated material pictures and imagery in these in these objects And also you would have linguists who are interested in the languages that they carry So we would have to branch or consider the interests of these different let's say types of researchers as well if you want to be a successful recommended system
Kyle Polich : For listeners who haven't yet visited and we'll put a link in the show notes to monasterium.net is that correct
Florian Atzenhofer : Yes that's
Kyle Polich : correct Could you share some details on what they'll find there and maybe what your contributions are
Florian Atzenhofer : So Monasterium.net has been around for more than 20 years It was created in the heart of Austria in Lower Austria by a set of monasteries who just wanted to provide a better way for interested researchers or just any other person to look at their very nice material So they used the traction of let's say the web at that time and the expertise of some
Florian Atzenhofer : IT people and came up with the very first version and the basic version that is a digital archive for charters It started small and over the years it grew up to be the largest online collection of these materials The project DDip which I'm currently working in and in conjunction to which my PhD is done rebuilds currently this platform
Florian Atzenhofer : And we're up to releasing the next version towards the end of the year So when you visit monasterium.net currently you will see the old version and there will not be a recommended system yet implemented And also not any of the new fancy semantic search features Those will be available by the end of the year
Kyle Polich : Can you maybe er tease a little bit about the user experience how will um someone who's coming to the site once those features roll out how will the experience evolve for them
Florian Atzenhofer : Well we differentiate between uh an exploration page and a more controlled way of looking for let's say similar items because we know that we need to differentiate as we discussed earlier between more expert users and more novice users
Florian Atzenhofer : And we found the best way to enable this is to separate the functionalities a little bit two components and this has results on the user experience which means if people are novices um they are more guided let's say or probably more prone to just using the exploration page
Florian Atzenhofer : They will be able to select different profiles that we have preconfigured and which are based on priorities of research or looking up items that we have found are important based on just our experience with working with these different people with the different users and the different stakeholders
Florian Atzenhofer : And then there will be other features such as selecting individual items or groups of items and building new sets of data based on that So you could have one or more let's say chart or baskets and they're called baskets and based on these sub-selections you can exploit them to find more similar ones The search is more based on re-ranking and less of a recommendation but rather a controlled exploration
Kyle Polich : Could we zoom in a little bit there I think most listeners who have a surface level understanding of recommender systems will know things like the a priori algorithm or user item similarity like we're matching maybe Jaccard distance and these sorts of things Are any of those techniques useful or do you have to adopt different methodologies
Florian Atzenhofer : In essence for a recommender system the most important thing that we have to consider here is that we have this sparse matrix of user and item interactions And we have a very very cold starting problem here which means that we leverage a lot the knowledge base behind this data set and the content of this data set And what this means for those that are not familiar with it is that we leverage different modalities of the charter data
Florian Atzenhofer : And we embed them in different spaces In addition to that we make it so you can combine them and also weight them a little bit So for instance we know that one major use case is based on finding uh new or similar items just based on the text content of one charger assuming here that we do have a transcription if it is not created automatically by our HDR pipeline which is the automatic text recognition
Florian Atzenhofer : If that's the case then a similarity search based on text could be very beneficial for them And this modality of text similarity could be of interest and this can be combined with other modalities such as a subspace selection of imagery For instance if people would be interested in a very specific imagery of for instance animals that are depicted in these charters and there are several of them they would weight that feature more
Florian Atzenhofer : Depending on how you prepare these um embedded spaces you can allow the user to exploit them Given that we are a project and uh we have a specific timeline we do make a selection of features or feature spaces so a selection of modalities that we find most important and give the power to the user to weigh them a bit
Florian Atzenhofer : This means that they can exploit these modalities and by this influence which things are being recommended to them And the side effect of this is that we can then fill this interaction matrix and make it easier for later users to use the whole system
Kyle Polich : Am I correct in saying that a charter is the item that you're ultimately gonna recommend
Florian Atzenhofer : Yes exactly At the very heart of it actually we focus on images so scans of charters which are usually very high quality And based on our pipelines we create let's see the text and based on this pipeline of the combination of a scan plus text recognition plus metadata that we either get from an edition or the archive directly we create a very
Florian Atzenhofer : Structured document which is in an XML encoding called the Charter Encoding Initiative This is an homage to the Text Encoding Initiative and has also been born 20 years ago and it it's still being used not just by our system
Florian Atzenhofer : And through this uh serialization we use one document and that is also one interpretation of such a manuscript in order to recommend it as an item So yes one charter is one item
Kyle Polich : So uh a charter it sounds like it's frequently an image but there could be metadata about it maybe you've transcribed something and then perhaps even that got translated I don't know how it works but metadata like the uh the date it was created or where it was found these sorts of things What's the field of metadata available to you
Florian Atzenhofer : So it depends on which pipelines have been applied to this specific original data We assume that we either have an XML serialization already provided with metadata from the archive because we're also working a lot with material that has already been digitized or has been around for some time uh or we get just the scans of images and apply different pipelines on it And depending on which pipelines we have different metadata fields available and they can then be embedded
Florian Atzenhofer : So for instance we can assume that we have for all of them some form of uh automatic text recognition Also in connection to this we have abstracts and they are called regesta in the proper terminology So essentially these are short summaries about what the charter is about And this is really important because charters can be very very lengthy
Florian Atzenhofer : And also in the past scholars of the 18th 19th century have employed writing abstracts or short forms of it just to be able to sort them better And we can leverage this tradition as well uh for similarity search and for recommendation
Florian Atzenhofer : In addition to a text an image an abstract we have much more information and this is in connection to the tradition and the discipline that is dedicated to the research of charters and this is called diplomatics And the field of diplomatics has come up with some features that are important such as who is the issuer And in connection to this
Florian Atzenhofer : To NLP applications we can leverage named entity recognition and the proper serialization for this This is not so easy because we're talking about historical material here We can leverage named entity recognition and a semantic labeling For instance this person is the issuer and the other person is the receiver and we can leverage this uh technology to create new embeddings as well And we can make them either implicit or explicit
Florian Atzenhofer : Depending on the use case depending on the downstream task
Kyle Polich : Makes it easy quick and safe to remove your personal data online at a time when surveillance and data breaches are common enough to make everyone vulnerable DeleteM isn't just a one-time service Delete Me is always working for you constantly monitoring and removing your personal information that you don't want on the internet Their team of privacy experts will remove your data from hundreds of data broker websites giving you comprehensive coverage where it matters most
Kyle Polich : As someone who values data privacy I've been alarmed by how my information keeps reappearing on broker sites even after I've manually requested removal That's why I appreciate DeleteM's ongoing service They don't just remove your data once they keep monitoring and removing it regularly saving me countless hours of frustration
Kyle Polich : Removing your data will help you protect yourself from potential identity theft and unwanted exposure Take control of your data and keep your private life private by signing up for Delete Me Now at a special discount for our listeners Get 20% off your Delete Me plan when you go to joindeleteme.com/data and use the promo code data at checkout The only way you get the 20% off is to go to joindeleteme.com/data
Kyle Polich : Enter code data at checkout That's join deleteme.com/datacodedata
Kyle Polich : And do you have any thoughts on how you'll measure the success of the system I guess ultimately you want people to find novel recommendations but how will you know if you're achieving that or to the degree you're achieving it
Florian Atzenhofer : Yes I mean there's very classic approaches to the evaluation and there are different metrics and there are things like uh root mean square deviation or error And you could look at metrics that are more important for aspects like fairness where you essentially look at the propensity or the amount of of charters
Florian Atzenhofer : Based on a ranked list which have been liked or preferred by one group of users
Florian Atzenhofer : Versus another So if you look at multi-stakeholder recommendation and this is important here we have to look at the overlap of preferred items versus these preferred items for a specific subset
Florian Atzenhofer : And we are only slowly figuring out what the best way of evaluating the recommendation system is We do have evaluation procedures for every other pipeline that goes before the recommendation but for um coming up with a full evaluation scheme uh we're not quite there yet This is in the works And this is also where my uh PhD or parts of my PhD come into play And what I've done over the past few
Florian Atzenhofer : Weeks and months actually is I've talked a lot with different stakeholders and I've done structured interviews and focus groups with several of them Uh there was a selection of 25 experts from a pool of over 70 plus
Florian Atzenhofer : And I've talked with upstream provider system consumer and downstream stakeholders and that would include a whole range of archivists specialists the developers also of of this system or rather similar systems but also the researcher educator students that use this material and use this platform for teaching And also lastly to produce scientific output And that's where also the downstream stakeholders come into play
Florian Atzenhofer : Because at the end of the day what is being shown in such a platform what users look at and as such what is being recommended to them could end up in a published article could end up being edited for school books even So essentially we need to consider different perspectives on what is being recommended
Florian Atzenhofer : And this can be leveraged for evaluation as well and I'm convinced that this is necessary for a public and less commercial domain
Kyle Polich : I'm curious if you got any specific insights from those user interview sessions I've been on a few situations like that in projects I've worked on and I've always found it very insightful to directly connect with my users and hear their needs
Florian Atzenhofer : Definitely definitely Many findings and I had the chance to also put this into a paper But what we found is that it's really not easy at all to structure the let's say the preferred values associated with the recommendation And we came up with what we call the research funnel or alignment to the research funnel where we differentiate between different phases of such an interaction between user and item
Florian Atzenhofer : So this begins with the discovery stage Yeah this is just about the exploration the first contact so to say
Florian Atzenhofer : It goes over to an interaction stage where the users uh let's say are a bit more cautious and hesitant about how this item might be useful to them in that very moment or even later
Florian Atzenhofer : Then it goes over to the integration stage where this question is already being settled and users think more deeply about the item For instance taking one item plus comparing it to another And we came up with also the fourth that is the impact stage Here things come into play that ask how useful is this for further exploration of sources how useful is this for actual publications
Florian Atzenhofer : And how important is this to other stakeholders
Florian Atzenhofer : We propose based on these stages different let's say metric directions This is not a full evaluation framework but rather it is uh a pathway towards it So we call them um evaluation setups
Florian Atzenhofer : And we propose different metrics such as research path quality This is a subjective measure of the user how it integrates with their specific research journey And we know that these researchers do have very individual journeys at times Sometimes you don't feel like that um sometimes you do
Florian Atzenhofer : And what we also found is that a lot of the stakeholders they are not only interested in explainability but they are also interested in the biases prevalent in these systems So what we found is that essentially all of them somehow find that
Florian Atzenhofer : We should consider collection representation as one measure that is we should be able to consider that one overly represented archive does not overshadow others And with this risk that some things some charters are just lost and
Florian Atzenhofer : Being behind in being lost behind in the shelves being put away and not seeing the daylight anymore So the digital daylight so to say This means collection representation is also important to them
Florian Atzenhofer : Yeah and we proposed several other metric directions which we want to explore And as soon as we have uh operationalized them as soon as we um formalized them um and implemented them we can also evaluate
Kyle Polich : So many years ago I saw a presentation by someone I believe they worked at Amazon.com uh on the recommended system team and they said
Kyle Polich : The recommender will be complete when you visit Amazon and there's a single product on the homepage and you immediately buy it which I hope they meant as a joke because it's sort of a silly but you know an extreme version of what's the goal So I guess my question for you is what's the goal Discovery's a big part of it like if I'm exploring these things and I'm a researcher a little bit of serendipity would be good learn things I didn't know I was looking for but I also need to get my research paper done
Kyle Polich : What uh how do you think about the best experience to create for a user in terms of what they're finding maybe versus discovering or getting right to it
Florian Atzenhofer : Well certainly uh just based on the nature of this domain serendipity is very important We're talking about research here and the context of cultural heritage So it's about finding stuff and uh finding some form of identification with it And this was also reflected in the focus group interviews that I've done for this past study
Florian Atzenhofer : Uh and it was quite interesting because stakeholders interpreted serendipity in different ways I've had one upstream stakeholder that it was an archivist uh and they argued the most beautiful finds you make are the findings that you would not expect OK And then they go on and you would rather find them by equal representation
Florian Atzenhofer : Which is quite interesting because you would rather have a notion that
Florian Atzenhofer : Disentangling the system from biases would not have an effect on the serendipity
Florian Atzenhofer : And another consumer stakeholder that was a developer I think No that was a a specialist
Florian Atzenhofer : They emphasized that
Florian Atzenhofer : They think that we need to constantly be challenged so that we don't pursue the wrong and where we think we're going So there are several notions of being challenged while also being able to trust the system
Florian Atzenhofer : And the consensus was also that at the end of the day this platform and this recommendation system should not be a co-author in a publication but at least a specific version of the recommended system should be acknowledged in a dissemination in a scientific output So when people have used them for research they should acknowledge that they did so and also which specific version
Florian Atzenhofer : Which is kind of hard to grasp because usually we assume that a recommendation scenario is very subjective So the question is does it really make sense to let's say acknowledge the system where it was rather a very specific and individual situation of a single user
Florian Atzenhofer : I'm not quite sure yet
Kyle Polich : That's very interesting yeah I've worked with some companies who will store a uh unique identifier telling you exactly which version of their software in the Git repository is running at the time that some observation was made You could make some similar approach where you give your users a receipt with an ID but I don't know what they do with that ID next although
Kyle Polich : it's interesting
Florian Atzenhofer : Yes exactly That's a very good point We're also including let's say the DUI so the builds of the specific software stack at that time and they very well can include that for any output uh or any citation But still the question is whether the users have really understood it and also if the user base can generally understand let's say the tech stack and the technology behind it
Florian Atzenhofer : Because most of the stakeholders lamented that they find the technology very interesting and they can imagine it is extremely useful to digital archives this one specifically but they also generalize to cultural heritage in general But they lamented as I said
Florian Atzenhofer : That they would like to understand the system as much as possible while it should be most effective and also carry explainability So we believe there must be some trade-off and the most easy form uh as you've put earlier would be just a single item recommendation and nothing else which of course takes away all the agency all the understanding process that scholars said to desire
Kyle Polich : Well tech similarity is something that has evolved quite a bit or your options have grown at least let's say over the last two decades Sometimes something simple like TFIDF will do the job for you and that's OK but we've also got you know uh neural network based things and embeddings and stuff like that What was the methodology that worked for you
Florian Atzenhofer : That's a very good question My thesis itself the work I've done was both theoretical and practical I've had a very very thorough introduction into different aspects into what I call formulacity or what is called formulacity That is essentially the formulaic language aspect that is prevalent in charters and that's very crucial
Florian Atzenhofer : And I showcased in my thesis theoretically and practically exactly that you have different ways of looking at similarity and similarity means different things to different folk Um and this is a let's say a a critical topic to this day in our team and uh to uh the associates Um
Florian Atzenhofer : It is really hard to measure it and you it it seems that you have to show them a lot especially the non-technical people what can be understood by a specific kind of text similarity As you've put earlier um
Florian Atzenhofer : It's true there are more simple measures but there are also more complex ones And at the end of the day it really depends on the downstream tasks And this means that we are bound to these tasks and they require us to to embed our data in a different way For instance there of course you can take TFIDF and this is rather still rather effective It is also the case for historical data it seems and this has
Florian Atzenhofer : Uh several reasons also due to the peculiarity of certain languages or certain terminology or certain regions And then you can go over to more let's say more sophisticated but still still very understandable embeddings based on our features For instance just skip grams or leaving out grams which means uh which is a a similar form to conk grams
Florian Atzenhofer : And what you do is you have certain phrases you leave words out and you still consider them to be part of a specific phrase or you also consider order of these words and still subsume it under the same group and then you do the embeddings
Kyle Polich : Well my hope for you as the recommender system rolls out is that you become a victim of your own success that everyone's very happy with it and they're demanding a little more hey how do we take this to the next level What's the next
Kyle Polich : level
Florian Atzenhofer : Uh well certainly I hope that the platform that we've built and renewed grows and grows and grows and uh things are looking very good
Florian Atzenhofer : Our project is also engaged in getting more data into the system It's becoming more let's say international We are filling the gaps that we have geographically in Europe such as more towards western of Europe and towards the north And actually we're talking about not just uh a handful of charters that is more in the system but tens and hundreds of thousands of charters more
Florian Atzenhofer : And further down the road we could imagine that
Florian Atzenhofer : Yeah recommendation systems as a means of exploration are being picked up by not just um our digital archives but also other digital archives that contain cultural heritage material
Florian Atzenhofer : Um we are convinced that or rather it is a fact that a lot of repositories that are currently concerned just with the preservation of already digitized material are silos They are yeah they are bunkers They are really closed and it's sometimes hard to get the data out without just flattening them and making them unusable I hope that
Florian Atzenhofer : Exploration through recommended systems will catch on with other archives and we could also if we go down one step further we could also imagine recommendation on a more aggregated level assuming that the data is not too flat So if
Florian Atzenhofer : Think about aggregation services be it on a national level or for instance on a European level such as with the Europa we will definitely see in the next let's say 2 years that exploration for recommendation will catch on and will be implemented
Kyle Polich : Well my personal bias is always towards like open things open source open data maybe a federated system open standards open protocols things like that Um you're free to disagree with me but uh either way do you have a sense of the state of openness and standards in digital humanities like uh how advanced is that and uh would you like to see it evolve in any
Kyle Polich : way
Florian Atzenhofer : I think you mentioned it especially in connection to the point I made earlier about the repositories but in general software stacks that are being used as repositories are mostly open source already but they're also changing technologies and it depends on
Florian Atzenhofer : For instance the institutional support also which software is being chosen as a repository software and whether this is a closed-source or um closed-source software or not So the researchers or the technologies uh technologists working in digital humanities or cultural heritage lam or repositories they don't have the sole say in designing this but they definitely favor open-source solutions and if they had the chance they would always go for open source
Florian Atzenhofer : I see the digital humanities community as a very open one and they have always been and they showcased this uh over many decades now that they can very easily mutually influence with other disciplines uh be inspired by them
Florian Atzenhofer : And this is also the case with let's say the latest traction it gathered from computer science applications being more mainstream also LLMs being taken up in their research and just showing people that uh you can do a lot with them
Kyle Polich : I know you mentioned you're currently involved in your PhD and we talked a little bit about er where you're headed with it Could you remind listeners er what you're doing and I guess what your next steps
Kyle Polich : are
Florian Atzenhofer : So I'm a 2nd-year PhD student and my progress is quite good I've already chosen the the specific fields or the specific subtopics of my PhD So these are 3 strands and I'm focusing on the multi-stakeholder perspectives of charter recommendation as well as the information seeking behavior of the specific users And third I'm focusing on the multi-modal similarity
Florian Atzenhofer : Um on different multi-modal similarity approaches that you can leverage based on cultural heritage data So currently I'm focusing on the multi-stakeholder problem uh on its evaluation I'm presenting a paper on this at the Recommended Systems conference in Prague which is in September
Florian Atzenhofer : And also I would be part of the PhD symposium where I'm again presenting and discussing my ideas and whether I'm on the right track here
Florian Atzenhofer : And later down the road I will do some more literature research and try to find similarities and dissimilarities between the information seeking behaviors of different recommended system user groups And I want to find out how from that perspective what I could improve in my application domain Further down the road or rather in parallel
Florian Atzenhofer : Since this is also very important in our project we're focusing a lot on multimodality and how we can leverage this and how we can use multimodal representation to give power to users in doing search And we really believe in this So that's where I'm heading currently
Kyle Polich : Well exciting stuff A lot of areas I hope we can follow up on I I imagine there'll be some interesting progress along the way to look at
Florian Atzenhofer : Yes I hope so
Kyle Polich : And is there anywhere listeners can follow you online
Florian Atzenhofer : I've put a Twitter handle in the description I think I'm not very active in it but you can still follow the project DD
Florian Atzenhofer : That I'm still in and any other follow-up project on that one and the URL would be dip.eu
Kyle Polich : Very cool We'll have a link in the show note as well Well Florian thank you so much for taking the time to come on and
Kyle Polich : share your work
Florian Atzenhofer : Thanks again for having me It was fun talking to you
Kyle Polich : You as well