Interpretable Video Game Recommendation

Abstract

Recent work has shown that sparse autoencoders (SAEs) can recover interpretable concepts from latent representations of deep learning models, but their applicability to recommender systems remains largely unexplored. This thesis investigates whether SAEs can extract interpretable, monosemantic concepts from game recommendation embed- dings and whether these concepts support transparent recommendation steering. Using a Steam game dataset, two collaborative-filtering backbones (Modified ALS and ELSA) are first evaluated under a chronological dataset split. Four SAE variants, including a proposed Tag-Supervised SAE, are then compared over a comprehensive hyperparameter grid search with and without Top-K sparsity, assessing reconstruction quality, concept interpretability, monosemanticity, and recommendation retention. Finally, the learned concepts are integrated into a web application supporting tag-based recommendation steering. The experiments show that Top-K sparsity is essential for learning interpretable concepts, while differences between SAE architectures are comparatively modest. The quality of the learned concepts depends strongly on the underlying embedding repre- sentation. As backbones, Modified ALS yields substantially cleaner and more steerable concepts than ELSA despite weaker recommendation...