Model-Agnostic Explainability for Sequential Recommendation: A Shapley-Based Framework With Multidimensional Evaluation

Abstract

Sequential recommender systems rely on users’ previous interactions to predict what they may prefer next. Although recent models can capture complex patterns in interaction histories, understanding which past interactions actually contribute to a recommendation remains difficult, particularly when the models use different internal architectures. This work addresses this issue through a model-agnostic framework for temporal explanation of sequential recommendations. The framework uses a common prediction interface to apply the same perturbation-based attribution procedure to four recommendation models: GRU4Rec, MFRec, BERT4Rec-lite, and TransformerRec. Experiments were conducted on the MovieLens-100K dataset, with explanations evaluated using faithfulness, robustness, counterfactual stability, k-sufficiency, attribution concentration, and mask-versus-shuffle consistency. The experimental results show that the attribution patterns differ across the recommendation models. GRU4Rec places greater importance on recent interactions, whereas MFRec produces a more concentrated attribution profile. BERT4Rec-lite and TransformerRec distribute attribution across a broader portion of the interaction history. The faithfulness analysis also shows that removing highly attributed interactions produces measurable changes in recommendation scores. For the representative GRU4Rec and TransformerRec cases, the reported AUC-gap values are 0.0709 and 0.0078, respectively. Across the broader evaluation, the framework records a counterfactual stability of 0.5527, a genre-attribution Gini coefficient of 0.4028, and a median k-sufficiency of 2. These results suggest that a common post-hoc explanation procedure can be used to examine how different sequential recommendation architectures make use of historical interactions while preserving differences in their attribution behaviour.