Spectral Biclustering Boosts Explainability in Recommender Systems
New method cuts retraining costs for explainable recommendations by grouping users and items.
Explainability in recommender systems is crucial for trust but existing post-hoc methods like observation-level deletion diagnostics are computationally prohibitive. This paper from Salmeron and Arévalo proposes a block-deletion approach using spectral biclustering to group similar users and items, then removes entire interaction blocks. This dramatically cuts the number of retraining runs while still providing counterfactual explanations at the level of user segments, item groups, and their interactions.
Tested on MovieLens and Amazon datasets with SVD and Neural Collaborative Filtering, the framework reveals that top-ranked recommendations are often more sensitive to specific interaction blocks than lower-ranked ones, with some blocks acting as supporting evidence and others harming quality. The analysis also uncovers heterogeneous sensitivity across user segments, offering diagnostic insights invisible to standard metrics. The method is model-agnostic but its explanations depend on the chosen block structure, highlighting a trade-off between scalability and granularity.
- Uses spectral biclustering to group users and items, enabling block-level deletion instead of individual removal.
- Reduces retraining procedures significantly compared to fine-grained deletion diagnostics.
- Found that top-ranked recommendations are more sensitive to specific interaction blocks, with some blocks having detrimental effects.
Why It Matters
Makes explainable AI practical for large-scale recommender systems, boosting transparency and trust without prohibitive compute costs.