From Explanation to Intervention: An Audit-Debug-Refine Pipeline for Sequential Recommendation

Abstract

Explainability methods for sequential recommendation are structurally passive: they describe model behaviour without influencing training. This paper closes that loop by deploying a model-agnostic Shapley-based framework as an active auditing instrument within a three-stage audit-debug-refine pipeline. The audit mechanism is formalised as a cooperative game G = (T, v) over interaction timesteps, where the characteristic function v : 2T → ℝ evaluates model confidence over item coalitions. We prove that exact Shapley computation for this game is #P-complete (Proposition 1) and derive that the Monte Carlo estimator achieves ε-approximation in O(T log(1/δ)/ε2) samples with probability 1−δ (Proposition 2). Applying the pipeline to GRU4Rec on MovieLens 100K identifies 236 explanation-unreliable users (25%). Population-level diagnosis rules out sequence sparsity (mean sequence length: 19.97 vs 19.96 for flagged vs unflagged users) and identifies low model confidence (0.012 vs 0.037, three-fold difference) as the root cause. Three targeted interventions — BPR Max pairwise ranking loss, popularity-weighted hard negative sampling, and entropy-based output regularisation — are applied in direct response to this diagnosis. The combined model reduces flagged users by 53% (95% CI: 48.1%–57.9%, p < 0.001), improves mean faithfulness from −0.056 to +0.133, and raises NDCG@10 from 0.4382 to 0.4543. A central empirical finding emerges: faithfulness and robustness behave as competing objectives when training interventions shift the model's confidence distribution — a tension that existing XAI evaluation benchmarks do not capture — quantified by the normalised faithfulness–robustness ratio FR_n.