In episode 29 of Recsperts, I welcome Craig Macdonald, Professor of Information Retrieval at the University of Glasgow, and Aleksandr “Sasha” Petrov, PhD researcher and former applied scientist at Amazon. Together, we dive deep into sequential recommender systems and the growing role of transformer models such as SASRec and BERT4Rec.
Our conversation begins with their influential replicability study of BERT4Rec, which revealed inconsistencies in reported results and highlighted the importance of training objectives over architecture tweaks. From there, Craig and Sasha guide us through their award-winning research on making transformers for sequential recommendation with large corpora both more effective and more efficient. We discuss how recency sampling (RSS) reduces training times dramatically, and how gSASRec overcomes the problem of overconfidence in models trained with negative sampling. By generalizing the sigmoid function (gBCE), they were able to reconcile cross-entropy–based optimization results with negative sampling, matching the effectiveness of softmax approaches while keeping training scalable for large corpora.
We also explore RecJPQ, their recent work on joint product quantization for item embeddings. This approach makes transformer-based sequential recommenders substantially faster at inference and far more memory-efficient for embeddings—while sometimes even improving effectiveness thanks to regularization effects. Towards the end, Craig and Sasha share their perspective on generative approaches like GPTRec, the promises and limits of large language models in recommendation, and what challenges remain for the future of sequential recommender systems.
Enjoy this enriching episode of RECSPERTS – Recommender Systems Experts.
Don’t forget to follow the podcast and please leave a review.
Papers:
General Links:
Disclaimer:
Craig holds concurrent appointments as a Professor of Information Retrieval at University of Glasgow and as an Amazon Scholar. This podcast describes work performed at the University of Glasgow and is not associated with Amazon.
In episode 28 of Recsperts, I sit down with Robin Burke, professor of information science at the University of Colorado Boulder and a leading expert with over 30 years of experience in recommender systems. Together, we explore multistakeholder recommender systems, fairness, transparency, and the role of recommender systems in the age of evolving generative AI.
We begin by tracing the origins of recommender systems, traditionally built around user-centric models. However, Robin challenges this perspective, arguing that all recommender systems are inherently multistakeholder—serving not just consumers as the recipients of recommendations, but also content providers, platform operators, and other key players with partially competing interests. He explains why the common “Recommended for You” label is, at best, an oversimplification and how greater transparency is needed to show how stakeholder interests are balanced.
Our conversation also delves into practical approaches for handling multiple objectives, including reranking strategies versus integrated optimization. While embedding multistakeholder concerns directly into models may be ideal, reranking offers a more flexible and efficient alternative, reducing the need for frequent retraining.
Towards the end of our discussion, we explore post-userism and the impact of generative AI on recommendation systems. With AI-generated content on the rise, Robin raises a critical concern: if recommendation systems remain overly user-centric, generative content could marginalize human creators, diminishing their revenue streams.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 27 of Recsperts, we meet Alessandro Piscopo, Lead Data Scientist in Personalization and Search, and Duncan Walker, Principal Data Scientist in the iPlayer Recommendations Team, both from the BBC. We discuss how the BBC personalizes recommendations across different offerings like news or video and audio content recommendations. We learn about the core values for the oldest public service media organization and the collaboration with editors in that process.
The BBC once started with short video recommendations for BBC+ and nowadays has to consider recommendations across multiple domains: news, the iPlayer, BBC Sounds, BBC Bytesize, and more. With a reach of about 500M+ users who access services every week there is a huge potential. My guests discuss the challenges of aligning recommendations with public service values and the role of editors and constant exchange, alignment, and learning between the algorithmic and editorial lines of recommender systems.
We also discuss the potential of cross-domain recommendations to leverage the content across different products as well as the organizational setup of teams working on recommender systems at the BBC. We learn about skews in the data due to the nature of an online service that also has a linear offering with TV and radio services.
Towards the end, we also touch a bit on QUARE @ RecSys, which is the Workshop on Measuring the Quality of Explanations in Recommender Systems.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 26 of Recsperts, I speak with Sanne Vrijenhoek, a PhD candidate at the University of Amsterdam’s Institute for Information Law and the AI, Media & Democracy Lab. Sanne’s research explores diversity in recommender systems, particularly in the news domain, and its connection to democratic values and goals.
We dive into four of her papers, which focus on how diversity is conceptualized in news recommender systems. Sanne introduces us to five rank-aware divergence metrics for measuring normative diversity and explains why diversity evaluation shouldn’t be approached blindly—first, we need to clarify the underlying values. She also presents a normative framework for these metrics, linking them to different democratic theory perspectives. Beyond evaluation, we discuss how to optimize diversity in recommender systems and reflect on missed opportunities—such as the RecSys Challenge 2024, which could have gone beyond accuracy-chasing. Sanne also shares her recommendations for improving the challenge by incorporating objectives such as diversity.
During our conversation, Sanne shares insights on effectively communicating recommender systems research to non-technical audiences. To wrap up, we explore ideas for fostering a more diverse RecSys research community, integrating perspectives from multiple disciplines.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 25, we talk about the upcoming ACM Conference on Recommender Systems 2024 (RecSys) and welcome a former guest to geek about the conference.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 24 of Recsperts, I sit down with Amey Dharwadker, Machine Learning Engineering Manager at Facebook, to dive into the complexities of large-scale video recommendations. Amey, who leads the Video Recommendations Quality Ranking team at Facebook, sheds light on the intricate challenges of delivering personalized video feeds at scale. Our conversation covers content understanding, user interaction data, real-time signals, exploration, and evaluation techniques.
We kick off the episode by reflecting on the inaugural VideoRecSys workshop at RecSys 2023, setting the stage for a deeper discussion on Facebook’s approach to video recommendations. Amey walks us through the critical challenges they face, such as gathering reliable user feedback signals to avoid pitfalls like watchbait. With a vast and ever-growing corpus of billions of videos—millions of which are added each month—the cold start problem looms large. We explore how content understanding, user feedback aggregation, and exploration techniques help address this issue. Amey explains how engagement metrics like watch time, comments, and reactions are used to rank content, ensuring users receive meaningful and diverse video feeds.
A key highlight of the conversation is the importance of real-time personalization in fast-paced environments, such as short-form video platforms, where user preferences change quickly. Amey also emphasizes the value of cross-domain data in enriching user profiles and improving recommendations.
Towards the end, Amey shares his insights on leadership in machine learning teams, pointing out the characteristics of a great ML team.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 23 of Recsperts, we welcome Yashar Deldjoo, Assistant Professor at the Polytechnic University of Bari, Italy. Yashar's research on recommender systems includes multimodal approaches, multimedia recommender systems as well as trustworthiness and adversarial robustness, where he has published a lot of work. We discuss the evolution of generative models for recommender systems, modeling paradigms, scenarios as well as their evaluation, risks and harms.
We begin our interview with a reflection of Yashar's areas of recommender systems research so far. Starting with multimedia recsys, particularly video recommendations, Yashar covers his work around adversarial robustness and trustworthiness leading to the main topic for this episode: generative models for recommender systems. We learn about their aspects for improving beyond the (partially saturated) state of traditional recommender systems: improve effectiveness and efficiency for top-n recommendations, introduce interactivity beyond classical conversational recsys, provide personalized zero- or few-shot recommendations.
We learn about the modeling paradigms and as well about the scenarios for generative models which mainly differ by input and modelling approach: ID-based, text-based, and multimodal generative models. This is how we navigate the large field of acronyms leading us from VAEs and GANs to LLMs.
Towards the end of the episode, we also touch on the evaluation, opportunities, risks and harms of generative models for recommender systems. Yashar also provides us with an ample amount of references and upcoming events where people get the chance to know more about GenRecSys.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 22 of Recsperts, we welcome Prabhat Agarwal, Senior ML Engineer, and Aayush Mudgal, Staff ML Engineer, both from Pinterest, to the show. Prabhat works on recommendations and search systems at Pinterest, leading representation learning efforts. Aayush is responsible for ads ranking and privacy-aware conversion modeling. We discuss user and content modeling, short- vs. long-term objectives, evaluation as well as multi-task learning and touch on counterfactual evaluation as well.
In our interview, Prabhat guides us through the journey of continuous improvements of Pinterest's Homefeed personalization starting with techniques such as gradient boosting over two-tower models to DCN and transformers. We discuss how to capture users' short- and long-term preferences through multiple embeddings and the role of candidate generators for content diversification. Prabhat shares some details about position debiasing and the challenges to facilitate exploration.
With Aayush we get the chance to dive into the specifics of ads ranking at Pinterest and he helps us to better understand how multifaceted ads can be. We learn more about the pain of having too many models and the Pinterest's efforts to consolidate the model landscape to improve infrastructural costs, maintainability, and efficiency. Aayush also shares some insights about exploration and corresponding randomization in the context of ads and how user behavior is very different between different kinds of ads.
Both guests highlight the role of counterfactual evaluation and its impact for faster experimentation.
Towards the end of the episode, we also touch a bit on learnings from last year's RecSys challenge.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 21 of Recsperts, we welcome Martijn Willemsen, Associate Professor at the Jheronimus Academy of Data Science and Eindhoven University of Technology. Martijn's researches on interactive recommender systems which includes aspects of decision psychology and user-centric evaluation. We discuss how users gain control over recommendations, how to support their goals and needs as well as how the user-centric evaluation framework fits into all of this.
In our interview, Martijn outlines the reasons for providing users control over recommendations and how to holistically evaluate the satisfaction and usefulness of recommendations for users goals and needs. We discuss the psychology of decision making with respect to how well or not recommender systems support it. We also dive into music recommender systems and discuss how nudging users to explore new genres can work as well as how longitudinal studies in recommender systems research can advance insights.
Towards the end of the episode, Martijn and I also discuss some examples and the usefulness of enabling users to provide negative explicit feedback to the system.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 20 of Recsperts, we welcome Bram van den Akker, Senior Machine Learning Scientist at Booking.com. Bram's work focuses on bandit algorithms and counterfactual learning. He was one of the creators of the Practical Bandits tutorial at the World Wide Web conference. We talk about the role of bandit feedback in decision making systems and in specific for recommendations in the travel industry.
In our interview, Bram elaborates on bandit feedback and how it is used in practice. We discuss off-policy- and on-policy-bandits, and we learn that counterfactual evaluation is right for selecting the best model candidates for downstream A/B-testing, but not a replacement. We hear more about the practical challenges of bandit feedback, for example the difference between model scores and propensities, the role of stochasticity or the nitty-gritty details of reward signals. Bram also shares with us the challenges of recommendations in the travel domain, where he points out the sparsity of signals or the feedback delay.
At the end of the episode, we can both agree on a good example for a clickbait-heavy news service in our phones.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 19 of Recsperts, we welcome Himan Abdollahpouri who is an Applied Research Scientist for Personalization & Machine Learning at Spotify. We discuss the role of popularity bias in recommender systems which was the dissertation topic of Himan. We talk about multi-objective and multi-stakeholder recommender systems as well as the challenges of music and podcast streaming personalization at Spotify.
In our interview, Himan walks us through popularity bias as the main cause of unfair recommendations for multiple stakeholders. We discuss the consumer- and provider-side implications and how to evaluate popularity bias. Not the sheer existence of popularity bias is the major problem, but its propagation in various collaborative filtering algorithms. But we also learn how to counteract by debiasing the data, the model itself, or it's output. We also hear more about the relationship between multi-objective and multi-stakeholder recommender systems.
At the end of the episode, Himan also shares the influence of popularity bias in music and podcast streaming at Spotify as well as how calibration helps to better cater content to users' preferences.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 18 of Recsperts, we hear from Professor Sole Pera from Delft University of Technology. We discuss the use of recommender systems for non-traditional populations, with children in particular. Sole shares the specifics, surprises, and subtleties of her research on recommendations for children.
In our interview, Sole and I discuss use cases and domains which need particular attention with respect to non-traditional populations. Sole outlines some of the major challenges like lacking public datasets or multifaceted criteria for the suitability of recommendations. The highly dynamic needs and abilities of children pose proper user modeling as a crucial part in the design and development of recommender systems. We also touch on how children interact differently with recommender systems and learn that trust plays a major role here.
Towards the end of the episode, we revisit the different goals and stakeholders involved in recommendations for children, especially the role of parents. We close with an overview of the current research community.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 17 of Recsperts, we meet Miguel Fierro who is a Principal Data Science Manager at Microsoft and holds a PhD in robotics. We talk about the Microsoft recommenders repository with over 15k stars on GitHub and discuss the impact of LLMs on RecSys. Miguel also shares his view of the T-shaped data scientist.
In our interview, Miguel shares how he transitioned from robotics into personalization as well as how the Microsoft recommenders repository started. We learn more about the three key components: examples, library, and tests. With more than 900 tests and more than 30 different algorithms, this library demonstrates a huge effort of open-source contribution and maintenance. We hear more about the principles that made this effort possible and successful. Therefore, Miguels also shares the reasoning behind evidence-based design to put the users of microsoft-recommenders and their expectations first. We also discuss the impact that recent LLM-related innovations have on RecSys.
At the end of the episode, Miguel explains the T-shaped data professional as an advice to stay competitive and build a champion data team. We conclude with some remarks regarding the adoption and ethical challenges recommender systems pose and which need further attention.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Don't forget to follow the podcast and please leave a review
Papers:
General Links:
In episode 16 of Recsperts, we hear from Michael D. Ekstrand, Associate Professor at Boise State University, about fairness in recommender systems. We discuss why fairness matters and provide an overview of the multidimensional fairness-aware RecSys landscape. Furthermore, we talk about tradeoffs, methods and receive practical advice on how to get started with tackling unfairness.
In our discussion, Michael outlines the difference and similarity between fairness and bias. We discuss several stages at which biases can enter the system as well as how bias can indeed support mitigating unfairness. We also cover the perspectives of different stakeholders with respect to fairness. We also learn that measuring fairness depends on the specific fairness concern one is interested in and that solving fairness universally is highly unlikely.
Towards the end of the episode, we take a look at further challenges as well as how and where the upcoming RecSys 2023 provides a forum for those interested in fairness-aware recommender systems.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Papers:
General Links:
In episode 15 of Recsperts, we delve into podcast recommendations with senior data scientist, Mirza Klimenta. Mirza discusses his work on the ARD Audiothek, a public broadcaster of audio-on-demand content, where he is part of pub. Public Value Technologies, a subsidiary of the two regional public broadcasters BR and SWR.
We explore the use and potency of simple algorithms and ways to mitigate popularity bias in data and recommendations. We also cover collaborative filtering and various approaches for content-based podcast recommendations, drawing on Mirza's expertise in multidimensional scaling for graph drawings. Additionally, Mirza sheds light on the responsibility of a public broadcaster in providing diversified content recommendations.
Towards the end of the episode, Mirza shares personal insights on his side project of becoming a novelist. Tune in for an informative and engaging conversation.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Papers:
General Links:
In episode number 14 of Recsperts we talk to Daniel Svonava, CEO and Co-Founder of Superlinked, delivering user modeling infrastructure. In his former role he was a senior software engineer and tech lead at YouTube working on ad performance prediction and pricing.
We discuss the crucial role of user modeling for recommendations and discovery. Daniel presents two examples from YouTube’s ad performance forecasting to demonstrate the bandwidth of use cases for user modeling. We also discuss sources of information that fuel user models and additional personlization tasks that benefit from it like user onboarding. We learn that the tight combination of user modeling with (near) real-time updates is key to a sound personalized user experience.
Daniel also shares with us how Superlinked provides personalization as a service beyond ecommerce-centricity. Offering personalized recommendations of items and people across various industries and use cases is what sets Superlinked apart. In the end, we also touch on the major general challenge of the RecSys community which is rebranding in order to establish a more positive image of the field.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Chapters:
Papers:
General Links:
This episode of Recsperts features Justin Basilico who is director of research and engineering at Netflix. Justin leads the team that is in charge of creating a personalized homepage. We learn more about the evolution of the Netflix recommender system from rating prediction to using deep learning, contextual multi-armed bandits and reinforcement learning to perform personalized page construction. Deep content understanding drives the creation of useful groupings of videos to be shown in a personalized homepage.
Justin and I discuss the misalignment of metrics as just one out of many elements that is making personalization still “super hard”. We hear more about the journey of deep learning for recommender systems where real usefulness comes from taking advantage of the variety of data besides pure user-item interactions, i.e. histories, content, and context. We also briefly touch on RecSysOps for detecting, predicting, diagnosing and resolving issues in a large-scale recommender systems and how it helps to alleviate item cold-start.
In the end of this episode, we talk about the company culture at Netflix. Key elements are freedom and responsibility as well as providing context instead of exerting control. We hear that being really comfortable with feedback is important for high-performance people and teams.
Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.
Chapters:
Papers:
General Links: