Highlights of RecSys 2020
RecSys goes virtual
Criteo had 10 attendees at RecSys 2020 which despite not being in “the marvelous city” proved to be as interesting in virtual mode as it has been in person. Also with online socials, caipirinha tutorials, games of “wolf” (organised by Criteo) and online karaoke, it was lots of fun too. The online experience is something that we are all getting used to, but the combination of Whova, Zoom and GatherTown did a good job of creating a satisfying experience. GatherTown allowed wandering around the posters and actually bumping in to people you knew and being able to chat. It wasn’t completely free of technical glitches but it was able to create some actual conference experience. It certainly augmented the zoom pre-recorded aspect and gave something fresh.
Keynotes and papers
Bias was a big theme at this year’s RecSys with one keynote on 4 Reasons Social Media Makes us Vulnerable to Manipulation by Filippo Menczer and another on Bias on Search and Recommender Systems by Ricardo Baeza-Yates. There was also a technical keynote: “You Really Get Me”: Conversational AI Agents That Can Truly Understand and Help Users.
There have been great advancements in natural language processing recently. A few papers have explored how this technology can be leveraged for recommendation tasks. The paper What does BERT know about books, movies and music? answered the question by showing that BERT knows rather a lot; and suggests that pertained neural networks may be effectively used as a component of recommender systems.
Doubly Robust Estimator for Ranking Metrics with Post-Click Conversions tackles an issue close to our heart at Criteo in terms of improving offline metrics in this case for post click conversions. It is great to see more attention being paid to the application of contextual bandits in the recommendation context which have low treatment effects and large action spaces making it different to the classical formulation.
In a similar spirit to last year’s best long paper the paper “Neural Collaborative Filtering vs. Matrix Factorization Revisited” argued that the use of dot products between embeddings often outperforms more complex methods using multi-layered perceptrons (MLP) to compute distances between items. They showed that despite the fact that the MLP can in theory implement the dot product it often performed worse than the simple matrix factorization approach.
This year, the best long paper went to Progressive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations. The paper tackles some of the difficulties in using multi-task learning in real systems, proposes a new architecture to overcome them and provides some convincing experiments and actual A/B test results, something really nice to see in a paper!
The best short paper went to ADER: Adaptively Distilled Exemplar Replay Towards Continual Learning for Session-based Recommendation. The paper presents a new algorithm for session based recommendation in a streaming setting and proposes new methods and losses to counteract catastrophic forgetting.
Tutorials and workshops
Criteo AI Lab gave a tutorial on Bayesian Value Based Recommendation which was well attended and was accompanied by notebooks demonstrating Bayesian recommendation using Stan and TensorFlow (notebooks 1,2,3). The tutorial gave the intuition and understanding and code behind our KDD 2020 paper (Video).
There were also tutorials on Feature Engineering for Recommender Systems (Video and Code); Counteracting Bias and Increasing Fairness in Search and Recommender Systems: (Video); Introduction to Bandits in Recommender Systems: (Video and Code); Adversarial Learning for Recommendation: Applications for Security and Generative Tasks — Concept to Code: (Video and Code). Having these resources online is a great resource for recommendation systems researchers!
The conferenced ended with workshops including the REVEAL workshop organized by Criteo’s Flavian Vasile and Olivier Koch as well as Netflix’s Maria Dimakopoulou and Yves Raimond, Microsoft’s Adith Swaminathan and Cornell’s Thorsten Joachims. Criteo had a strong presence with Otmane Sakhi presenting his work as a talk on Improving Offline contextual bandits with distributional robustness and Philomene Chagniot presenting a poster From Clicks to Conversions: Recommendation for long-term reward. There is too much good stuff to go into on REVEAL, so there will be a second blog post on it coming very soon.
All up it was a great experience and we are looking forward to RecSys 2021 (in person or virtual).
Like what you are reading? Check out our latest publications on Medium!
