Spectral Biclustering-Driven Scalability for Post-Hoc Explainability in Recommender Systems

This paper addresses the scalability bottleneck in post-hoc explainability for recommender systems. Traditional deletion diagnostics—which involve retraining a model after removing individual user-item interactions—are computationally prohibitive for large datasets. To solve this, the authors propose a block-deletion framework. Instead of analyzing individual observations, the system uses spectral biclustering to partition the user-item interaction matrix into coherent blocks. These blocks are then systematically removed, and the model is retrained to observe the impact on recommendation quality. This approach significantly reduces the number of required retraining procedures while maintaining a model-agnostic, counterfactual perspective.

The framework was evaluated using Singular Value Decomposition (SVD) and Neural Collaborative Filtering (NCF) on the MovieLens and Amazon datasets. The study reveals that top-ranked recommendations are generally more sensitive to specific interaction blocks than lower-ranked ones. Furthermore, the analysis identifies that certain interaction blocks act as supportive evidence for recommendations, while others have a detrimental effect. The researchers also observed that different user segments exhibit varying levels of sensitivity to block removal, indicating that reliance on localized interaction patterns is heterogeneous across the user base.

Standard recommendation metrics often fail to provide diagnostic insights into why a model makes specific predictions. By reformulating retraining-based explainability at the block level, this framework offers a practical way to audit complex models without needing surrogate models or gradient-based approximations. This provides practitioners with a scalable tool to identify bias, failure modes, and the specific interaction patterns that drive model behavior, which is essential for ensuring transparency and accountability in high-stakes AI deployments.