A Systematic Benchmark of Explainable Methods for Temporal Attribution in Sequential Recommendation Systems

Gradient-based explanation methods outperform alternatives: GradientSHAP and Integrated Gradients proved most reliable at identifying which past user interactions actually drive recommendations, outperforming attention-based and perturbation-based approaches across multiple recommendation systems.

Attention weights alone are misleading: The paper found that simply using a model's built-in attention mechanisms (which are often assumed to show importance) do not reliably explain which past interactions influenced a recommendation, highlighting a critical gap between what models attend to and what actually matters for their decisions.

Systematic testing reveals hidden biases: The benchmark uncovered that item popularity can confound attribution results—meaning popular items may appear more important than they actually are—suggesting explainability methods need to account for these statistical artifacts to provide trustworthy explanations to users.

No single method works universally: Different explanation methods showed varying performance across different recommendation architectures (CNN, Transformer, SASRec, BERT4Rec), indicating that organizations cannot rely on one-size-fits-all approaches when auditing their recommendation systems for transparency.

Robustness matters for real-world deployment: The study tested how explanation methods held up when recommendation systems encountered corrupted or unusual input data, revealing that some methods' explanations became unreliable under realistic conditions—an important consideration for systems users actually depend on.