A Systematic Benchmark of Explainable Methods for Temporal Attribution in Sequential Recommendation Systems
Akash Pandey, Kanisha Shah, Addrish Roy, Dwipam Katariya, Hongyangyang Shi, Amanda Ding, Kalanand Mishra, Pranab Mohanty
cs.LG
Quick verdict
A mostly convincing benchmark of 10 XAI methods for temporal attribution in sequential RecSys, weakened by 100-user sampling and internal inconsistencies.
Summary
The paper benchmarks ten post-hoc explainability methods—Integrated Gradients, GradientSHAP, DeepLIFT, DeepLIFT-SHAP, Grad-CAM, LIME, KernelSHAP, Attention Tracing, Grad-SAM, and the intrinsically interpretable TimeSliver—for temporal attribution in sequential recommendation. Its methodological contribution is a dual-model faithfulness metric: an attribution model trained on complete sequences supplies per-timestep scores, while an architecturally identical probe trained with stochastic masking (p=0.05) measures the probability drop when top-attributed timesteps are masked, keeping perturbed inputs in-distribution. Experiments cover CNN, vanilla Transformer, SASRec, BERT4Rec, and TimeSliver backbones on KuaiRand-1k (downsampled to 100 users, L=100/1000) and MovieLens-1M (L=200). Results (Table 1) show gradient methods and Grad-SAM reaching c_f < −0.9 in many settings, raw Attention Tracing inconsistent (anti-faithful +0.930 on the vanilla Transformer with KuaiRand), Grad-CAM unfaithful on CNN, and Grad-SAM degrading on long sequences (−0.308 at L=1000). Robustness degradation stays within 25% under noise and popularity bias is near zero. Main weaknesses: the 100-user KuaiRand subsample, unexplained numerical identity of DeepLIFT and DeepLIFT-SHAP, a contradiction between Appendix A.2 and Tables 1/3/5 on where SASRec/BERT4Rec were trained, an unstated uncertainty basis in Table 1, and validation of Model A attributions against Model B predictions. No code release is stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Gradient-based methods, particularly Integrated Gradients and GradientSHAP, provide the most faithful and robust temporal attributions for sequential recommenders across the tested settings.
- Raw attention weights are unreliable as explanations, while gradient-weighted attention (Grad-SAM) is faithful on shorter sequences but degrades on long horizons (c_f = -0.308 at L=1000) as softmax attention approaches uniformity.
- The dual-model masking metric keeps perturbed inputs in-distribution via a masking-trained probe, avoiding both the OOD pitfall of naive deletion and the cost of remove-and-retrain evaluation.
- Faithful methods show negligible popularity bias (c_p within [-0.146, +0.110] vs. random baseline 0.007), and their recency patterns reflect dataset-dependent task structure rather than position or frequency confounds.
Strengths
Practical, well-motivated faithfulness metric
The dual-model probe (Section 3, Fig. 1) sidesteps the out-of-distribution artifacts of naive masking and the cost of remove-and-retrain (ROAR): one reusable probe per backbone, trained with masking probability p=0.05 matched to evaluation-time masking.
Unusually broad systematic coverage
10 methods x 5 backbones x 3 dataset-length settings, with a random-attribution sanity baseline (Table 1: c_f between -0.071 and +0.064) and four complementary axes: faithfulness, robustness (Fig. 2), recency profiles (Fig. 3), and popularity bias (Table 6).
Nuanced, actionable findings on attention
Intrinsic interpretability included as reference
Above-average implementation transparency
Weaknesses
Data-integrity anomalies in headline tables
Cross-model validity of the metric is asserted, not tested
Very small user sample on KuaiRand
Thin statistical treatment
Mild overclaiming and shallow limitations section
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.27201, CARS@RecSys'26) claims the first systematic benchmark of post-hoc XAI methods for temporal attribution in sequential recommendation. It evaluates ten methods — Integrated Gradients, GradientSHAP, DeepLIFT, DeepLIFT-SHAP, Grad-CAM, LIME, KernelSHAP, Attention Tracing, Grad-SAM, and the intrinsically interpretable TimeSliver — across CNN, vanilla Transformer, SASRec, BERT4Rec, and TimeSliver backbones on KuaiRand-1k (binary long-view classification, L=100/1000) and MovieLens-1M (next-item, L=200). Three axes are scored: faithfulness (c_f), robustness under input noise (c_r), and attribution mechanics (recency profiles, popularity bias c_p). Central claims: gradient methods (IG, GradientSHAP) are most faithful and robust; raw attention is unreliable; Grad-SAM is faithful on short sequences but degrades for L > 300; faithful attributions show negligible popularity bias.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the dual-model masking faithfulness metric proposed in this paper?
It tests whether an XAI method's per-interaction importance scores truly drive a sequential recommender's prediction. Model A, trained on full sequences, produces attribution scores; a second identical model (Model B) trained with random 5% input masking acts as a probe: timesteps are masked in groups ordered by attribution, and the Pearson correlation between each group's attribution and the resulting drop in Model B's predicted probability measures faithfulness (more negative = more faithful). Matching training and evaluation masking keeps perturbed inputs in-distribution.
Which explainability method works best for sequential recommendation models?
According to this benchmark, gradient-based methods - Integrated Gradients and GradientSHAP - are the most faithful and robust overall, with faithfulness correlations often below -0.9. Grad-SAM (gradient-weighted attention) is also strong but degrades for sequences longer than roughly 500 steps. Raw attention tracing and Grad-CAM were unreliable or unfaithful in several settings.
Are attention weights reliable explanations in sequential recommenders?
Not on their own. The paper finds raw attention tracing is highly architecture- and dataset-dependent, even anti-faithful (+0.930) on a vanilla Transformer with KuaiRand. Reweighting attention by gradients (Grad-SAM) restores faithfulness on shorter sequences, but it weakens on long horizons as softmax attention converges toward uniform weights.
What datasets and models were used, and is code available?
KuaiRand-1k (binary long-view prediction, downsampled from 1,000 to 100 users, sequence lengths 100 and 1000) and MovieLens-1M (next-item prediction, length 200), with CNN, vanilla Transformer, SASRec, BERT4Rec, and TimeSliver backbones. Implementations use Captum and are described in detail, but the paper does not mention a public code release.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-6 appear only as axis-label fragments in the extracted text; visual curve shapes (robustness curves, recency profiles) could not be verified.
- Appendix A.1 (dataset details) is empty in the extraction.
- Table 6's Grad-SAM SASRec row shows dashes for the KuaiRand popularity values; the missing entries are unexplained in the text.
- Flagged anomalies (e.g., DeepLift/DeepLiftSHAP duplication) could in principle be extraction artifacts, but they recur consistently across two separate tables, so they are treated as likely genuine.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.