Disentanglement and Interpretability in Recommender Systems

30:33

A good recommendation can feel effortless, even when the machinery behind it is anything but.

In this episode, Kyle talks with Ervin Dervishaj, a third-year PhD student at the University of Copenhagen specializing in recommender systems and representation learning. Ervin shares his journey from discovering recommender systems during his master's degree to pursuing cutting-edge research in the field. The conversation explores representation learning—how computers build their own understanding of data rather than relying on hand-crafted features—and the critical concept of disentanglement, where different factors like price and size in product recommendations remain independent of each other. The discussion centers on Ervin's research paper examining the relationship between representation disentanglement and interpretability in recommender models. While his team found a strong positive correlation between disentanglement and interpretability, surprisingly they did not find consistent correlation between disentanglement and recommendation performance. Ervin explains that disentanglement acts as a regularizer, potentially sacrificing some performance for improved interpretability and user trust. The conversation also touches on the importance of reproducibility in research, the role of large language models in denoising user profiles, and how disentangled representations could give users more control over their recommendations.

Ervin Dervishaj

University of Copenhagen

Applied Scientist intern Amazon and PhD fellow in Recommender Systems @ University of Copenhagen