In a detailed review of RecSys 2020, Eugene Yan identifies several recurring themes that signal where recommender systems research is heading. One of the strongest threads is bias: papers at the conference examined how biases can enter the data and the model, and how they might be mitigated to make recommendations more fair and reliable.
Another notable trend is the growing use of sequential models and bandit algorithms. These approaches treat recommendation as a dynamic problem, accounting for the order of user interactions and the need to balance exploration with exploitation. The shift suggests that static, one-shot recommendation is giving way to more adaptive systems.
Yan also stresses the importance of robust offline evaluation and the challenges of deploying recommenders in production. The conference included work on making offline experiments more trustworthy and on understanding how systems behave 'in the wild,' where conditions differ from controlled benchmarks. Together, these themes point to a maturing field that is paying closer attention to the real-world consequences of algorithmic choices.