Data Skeptic Disentanglement and Interpretability in Recommender Systems
13 snips
Mar 10, 2026 Erwin Dervishai, a PhD student at the University of Copenhagen who studies representation learning and recommender systems. He explores what disentanglement means for learned embeddings. He discusses methods for interpreting embeddings, reproducibility challenges, trade-offs between interpretability and accuracy, using metadata and LLMs for denoising, and practical ideas for user control.
AI Snips
Chapters
Transcript
Episode notes
Representation Learning Replaces Manual Features
- Representation learning replaces manual feature engineering by letting models learn vectors that summarize inputs for downstream tasks.
- Erwin Dervishai contrasts unsupervised clustering-style learning with supervised learning where labels guide the internal representation formation.
Disentanglement Separates Independent Factors
- Disentanglement aims to make learned representation dimensions correspond to independent factors like size or price.
- Erwin gives a t-shirt example: changing size should not change price if factors are disentangled.
Use Quantitative Disentanglement Metrics
- Evaluate disentanglement quantitatively using established metrics like disentanglement and completeness instead of only qualitative inspection.
- The authors collected prior models and datasets and applied these metrics in a reproducibility study.