#29: Transformers for Recommender Systems with Craig Macdonald and Sasha Petrov

In episode 34 of Recsperts, I'm joined by Raghav Saboo, Staff Machine Learning Engineer at DoorDash and Tech Lead for Personalization and Search for New Verticals — groceries, convenience, retail, alcohol, pet supplies and more, beyond the original restaurant vertical. We discuss the particular challenges of personalized recommendation, ranking and search in quick commerce, recent trends in generative recommendations and the application of Semantic IDs to item ranking and query reformulation. Raghav's path into recommender systems started in chemical engineering before he moved into ML consulting, a Master's in Statistics, Machine Learning and Econometrics from Duke University, and building LLMs for new language launches on Amazon's Alexa AI, ahead of joining DoorDash. We start with the marketplace itself: DoorDash connects consumers, merchants and couriers, and growing it means balancing the interests of all three so that the platform stays healthy for everyone on it. Raghav walks me through how his team frames the consumer side around three pillars — familiarity (surfacing what a consumer already trusts), affordability (matching price sensitivity and timely deals) and novelty (introducing new items and categories without adding friction). From there we get into how DoorDash uses LLMs to build "memory blocks," structured natural-language representations of a consumer organized around semantic domains like dietary preference, pet ownership or trusted brands, and how these feed LLM-generated collections that get resolved into real items through embedding-based retrieval. We then turn to DoorDash's move to generative approaches, centered on Semantic IDs: hierarchical product identifiers learned through recursive clustering of item content embeddings, forming a taxonomy that captures attributes a human-built catalog structure might miss — as Raghav puts it, "within e-commerce, items really carry a lot of meaning." He walks me through two production use cases: replacing dozens of taxonomy-based dense features in the ranking model with Semantic ID n-gram aggregations while improving online metrics, and using Semantic IDs for query reformulation in search, letting the system traverse a learned hierarchy to refine or diversify a query. This connects to DoorDash's own paper on the topic and to a broader conversation about why search, recommendation and agentic ordering — DoorDash's own "Ask DoorDash" — are converging on a shared substrate of Semantic IDs and consumer memory, while today's app surfaces still need to grow more flexible for that convergence to feel seamless. We close with a preview of the RecSys 2026 tutorial "Recommender Systems in Delivery Platforms: Challenges, Solutions and Learnings," which Raghav is co-presenting with Wolt's Paavo Camps and myself, and his advice for navigating a field that reinvents itself every quarter: be honest about whether that pace suits you, use AI agents to filter what's worth your attention, and build the judgment to recognize dead ends early. Enjoy this enriching episode of RECSPERTS – Recommender Systems Experts. Don't forget to follow the podcast and please leave a review.

In episode 33 of Recsperts, I speak with Joseph A. Konstan, Distinguished McKnight University Professor and Distinguished University Teaching Professor at the University of Minnesota, co-founder of GroupLens, and the very first General Chair of the ACM RecSys conference back in 2007. This year he returns in that role as General Co-Chair of the 20th RecSys in Minneapolis. We talk about what actually makes a recommendation useful, why the field is more than a machine learning application, and how the community around RecSys came into being. We start with what Joe cares about most: usefulness. He recalls the supermarket thought experiment of printing "buy bananas and bread" on every shopping cart and explains why the early systems were valuable because they were wrong a lot. A recommender that took ten people like you and often got it wrong was also often surprising when it got it right, whereas today's systems are wrong far less and useless far more. As Joe puts it, "I don't care about prediction at all. I care about changing people's behavior." We discuss why optimizing for click-through in news reliably produces clickbait, why leave-one-out evaluation only makes sense if you assume the user already knew the right answer and had simply forgotten it, and why usefulness can never be read off a single metric but depends on the task, the context and the breadth of the user's intent. Sometimes the most useful thing is not the recommendation itself, but the stars, the reviews or the comparison table you put around it. From there we turn to the community itself. Joe traces it back to the spring of 1996 and the Berkeley Collaborative Filtering Workshop organized by Hal Varian and Paul Resnick, through a decade of scattered workshops at SIGIR, CSCW and CHI, to the first RecSys in 2007: a room on the Minnesota campus, over 120 people, and a substantial delegation from industry including Amazon. We discuss why the conference has always been heavily international and always at the intersection of research and practice, what surprised him most in 20 years (that we are still here and thriving), and the vision he pushed against the pull of becoming just another application of machine learning: a highly constrained, highly contextualized, multi-measure and often multi-stakeholder problem spanning algorithms, interfaces, data and business. This year the 20th RecSys comes home to Minneapolis, with Joe as General Co-Chair alongside George Karypis and Gediminas Adomavicius. We close on where things are heading. Joe's advice for newcomers is to immerse yourself in an application and find the real problems and opportunities there, rather than arriving enamored with a tool and treating every nail as something to hit. He picks up Karl Higley's point from the POPROX team that the research community keeps focusing on the model when the real object is the system: your algorithm is useless if it is not embedded in something that can deliver its results in a useful way. And he names the work he is most excited about — human decision-making and consumer psychology in the context of choice, the business school perspective that brings marketing and pricing into the picture, multi-sided marketplaces with their often invisible market maker, and the ethics of these systems as a practical lens on the long-term value our metrics still fail to capture.

In episode 32 of Recsperts, I’m joined by my colleague Sasha Fedintsev, Staff Applied Scientist at Wolt (DoorDash), working across personalization and ads, to unpack the realities of building large-scale recommender systems in food, grocery, and retail delivery. Together, we discuss the specifics of personalization in the delivery domain, and the models and ideas that power Wolt’s recommender system across 30+ markets - where theory quickly meets messy, high-stakes practice. We explore what makes this domain fundamentally different from traditional e-commerce: strong locality constraints, real-time context, and a heavy skew toward repurchasing behavior. Sasha explains how these factors break many textbook approaches - like standard collaborative filtering - and require creative adaptations such as clustering strategies and multi-stage ranking systems optimized for latency, all while respecting locality constraints. We also discuss the evolution of recommendation approaches over time - from classical collaborative filtering with ALS, to Neural Collaborative Filtering with BPR, and ultimately to transformer-based models for user sequence modeling and next-purchase prediction powering today’s venue ranking systems. We also touch on practical challenges such as evaluation in real-world systems, including A/B testing pitfalls and biases in logged data, as well as the complexity introduced by multi-surface experiences like discovery pages, vertical lists, and search. Beyond venues, we discuss why item-level recommendation is an order of magnitude harder - due to scale, context dependence, and availability constraints - and what this implies for future system design. Throughout the episode, Sasha provides a candid view on the evolving role of a Staff Applied Scientist - bridging research and production, setting scientific standards, and driving cross-team impact. Enjoy this enriching episode of RECSPERTS – Recommender Systems Experts. Don’t forget to follow the podcast and please leave a review.

In episode 31 of Recsperts, I sit down with Elisabeth Lex, Full Professor of Human-Computer Interfaces and Inclusive Technologies at Graz University of Technology and a leading researcher at the intersection of recommender systems, psychology, and human–computer interaction. Together, we explore how recommender systems can become truly human-centric by integrating cognitive, emotional, and personality-aware models into their design.

In episode 30 of Recsperts, I speak with Annelien Smets, Professor at Vrije Universiteit Brussel and Senior Researcher at imec-SMIT, about the value, perception, and practical design of serendipity in recommender systems. Annelien introduces her framework for understanding serendipity through intention, experience, and affordances, and explains the paradox of artificial serendipity - why it cannot be engineered, but only designed for.