RecSys 2020, Key Takeaways — Part 1

RecSys is a ACM conference on Recommender systems. This year, it was scheduled to be in Rio De Janeiro, Brazil from Sep 22 to 24, 2020. But due to COVID-19 pandemic, it was converted to a fully virtual conference.

I attended RecSys conference virtually last week. The video recordings of the paper presentations, tutorials and workshop are now available on Vimeo: ACM RecSys 2020 . There were many interesting papers, presented on a variety of topics like deep models for recommender systems, fairness in recommendations, filter bubbles and echo chambers, contextual bandits for online learning etc. Many industry leaders like Amazon, Netflix, Alibaba, Spotify, Tencent, etc participated in the conference and spoke about their works.

Here are some of my notes from the sessions that I attended in the conference. (Continued in Part 2)

KeyNote Speech: How social media manipulates people?

Filipo Menczer, Indiana University

The speaker spoke about some of the key problems in social media platforms, and presented results from experiments conducted by his group on Twitter:

• Echo Chambers: People tend to follow more of like minded people and unfollow people of other social inclinations. Information in support of one group, fake or credible, tends to spread more in some groups than the other. Fact checks often don’t propagate to the audience following the fake news. Algorithms have been able to slow this, but unable to eliminate it.
• Information overload/limited attention span: Even if people genuinely share good quality information, because of limited attention span, they often cannot assess the credibility of all the information on their feed. With their limited attention span, they tend to share what looks more credible to them. Hence, what trends may not always be high quality information. Distribution of fake news vs credible news being shared looks similar. Once a news attracts enough traction, people tend to share it without thinking about its credibility.
• Social Bots: Automated bots often share fake news, in a coordinated manner targeting few popular individuals. Their goal is to be noticed by some popular politicians, journalists etc, who might fall for their trap and re-share/retweet the fake news. Once, popular people share it, more of their followers do the same and the news spreads to a larger audience quickly.

Selected Long Papers

SSE-PT: Personalized Transformer

Image taken from Liwei Wu, et al: SSE-PT: Sequential Recommendation Via Personalized Transformer

Recommender Systems often look at a sequence of items and recommend the next most relevant item to the user. This is similar to NLP sentence completion. Transformers have been effective for this task in NLP. So, the same might be useful in Recommender Systems. Some of the past developments in NLP have similarly found applications in building recommender systems like RNNs (GRU4Rec), Self-Attention (SASRec) etc.

PURS: Unexpected recommendations for improving user satisfaction (Alibaba)

Images taken from Pan Li et al: PURS: Personalized Unexpected Recommender System for Improving User Satisfaction

To break echo chambers/filter bubbles and to prevent recommending users items that they have already bought. Make multiple clusters of interested items for a user in embedding space. For a new item, find distance from clusters and penalize distances too far or too close (using a function called unexpected activation with minima shifted slightly away from 0, eg: xe^-x). Using a extra regularization term with unexpected activation, injects relevant novel results to the recommender systems, improving user experience.

PLE: Progressive layered extraction for multi task learning (Tencent) (Best paper award)

This paper builds on top of Google’s MMoE: Multi-Gate Mixture of Experts (KDD’18). MMoE trains a single shared set of experts for multiple tasks, with gated combinations for each task. PLE introduces a dedicated expert for each task along with a set of shared experts for all tasks. The paper compares and contrasts their approach with other such architectures like Sluice, Cross-Stitch.

Ensuring Fairness in Group Recommendations by Rank-Sensitive Balancing of Relevance

This paper introduces a new notion of fairness called GFAR (Group Fairness-Aware Recommendation). In group recommendations, the system needs to recommend the same list of items to a group of users. Aggregating the relevance scores of individual users into one single score for each item may not be equally appropriate for all users in the group. GFAR makes sure the aggregations are such that, the top N items from the rankings are as much relevant as possible for all users in the group.

Making Neural Networks Interpretable (Deezer Music Recommendation)

Image taken from Darius Afchar, Romain Hennequin: Making Neural Networks Interpretable with Attribution: Application to Implicit Signals Prediction

Train a neural network which is equivalent of an ensemble of neural networks each on a smaller subset of features. The scores are finally combined with a weighted average. The weights indicate which subset of features are more important.

Recommendations as Graph Exploration

A larger set of recommendation systems approach can be modeled as a graph exploration problem. The data and relationships can be modeled as nodes and edges in the graph. The recommendations become a problem of inferring new edges in the graph. The authors define a path algebra model with operations edge AGGregate, path FUSE, path CONcate and implement the system on Neo4j graphDB.

Announcements

Microsoft open sourced their new recommendation dataset: MIND

Google open sourced Tensorflow Recommender System

Link to Part 2: RecSys 2020, Key takeaways — Part 2