PaperlayerView exact revision on arXiv

Source-linked AI summary

Fidelity-Aware Recommendation Explanations via Stochastic Path Integration

Oren Barkan, Yahlly Schein, Yehonatan Elisha, Veronika Bogina, Mikhail Baklanov, Noam Koenigstein

arXiv:2511.18047v1cs.IRcs.AIcs.LG

TL;DR

Fidelity—whether recommender explanations reflect models’ true reasoning—remains underexplored. SPINRec adapts path integration using sampled plausible user profiles, and evaluations across models, datasets, and counterfactual metrics report consistent gains over baselines.

  • Problem

    Explanation fidelity remains underexplored in recommendation, despite its importance for transparency, accountability, and assessing whether explanations reflect a model’s reasoning.

    Source: Introduction · p. 1

  • Method

    SPINRec adapts path integration to recommendation by sampling plausible user baselines from empirical data and selecting the explanation that maximizes fidelity.

    Source: Introduction · p. 1

  • Results

    Across models, datasets, and fidelity metrics, SPINRec consistently outperforms existing approaches, establishing a new benchmark for fidelity-aware recommendation explanations.

    Source: Conclusion · p. 6

  • Takeaways & Limitations

    SPINRec produces stable, high-fidelity explanations tailored to sparse, binary user–item data.

    Source: Conclusion · p. 6

Abstract

from arXiv · show

Explanation fidelity, which measures how accurately an explanation reflects a model's true reasoning, remains critically underexplored in recommender systems. We introduce SPINRec (Stochastic Path Integration for Neural Recommender Explanations), a model-agnostic approach that adapts path-integration techniques to the sparse and implicit nature of recommendation data. To overcome the limitations of prior methods, SPINRec employs stochastic baseline sampling: instead of integrating from a fixed or unrealistic baseline, it samples multiple plausible user profiles from the empirical data distribution and selects the most faithful attribution path. This design captures the influence of both observed and unobserved interactions, yielding more stable and personalized explanations. We conduct the most comprehensive fidelity evaluation to date across three models (MF, VAE, NCF), three datasets (ML1M, Yahoo! Music, Pinterest), and a suite of counterfactual metrics, including AUC-based perturbation curves and fixed-length diagnostics. SPINRec consistently outperforms all baselines, establishing a new benchmark for faithful explainability in recommendation. Code and evaluation tools are publicly available at https://github.com/DeltaLabTLV/SPINRec.

Introduction

Recommendation explanations often emphasize user-facing qualities, while fidelity—whether they reflect a model’s actual reasoning—remains underexplored. SPINRec adapts path integration to sparse recommendation data, and its evaluation reports a new state-of-the-art benchmark.

Motivation

  • Fidelity remains an underexplored dimension of explainable recommendation, despite its importance for assessing whether explanations reflect a recommender’s actual decision process.Without fidelity, explanations may appear plausible while failing to reveal the model’s reasoning.Source: Introduction · p. 1

Method

  • SPINRec samples plausible user profiles from the empirical data distribution and selects the explanation maximizing fidelity, rather than integrating from an unrealistic all-zero baseline.The approach addresses sparse binary interactions and captures signals from both observed and unobserved interactions.Source: Introduction · p. 1

Results

  • SPINRec establishes a new state-of-the-art benchmark for recommendation explainability, with ablations confirming contributions from path integration and stochastic baseline sampling.The evaluation spans MF, VAE, and NCF models; ML1M, Yahoo! Music, and Pinterest; and counterfactual fidelity metrics.Source: Introduction · p. 1

Contributions

  • The paper introduces SPINRec as the first adaptation of path-integration methods to recommender systems and presents fidelity-focused evaluation across models and datasets.The authors describe the work as groundwork for research on trustworthy, model-faithful explanations.Source: Introduction · pp. 1–2

Related Work

Prior recommender-explanation work includes model-specific and model-agnostic methods, but fidelity evaluation remains underexplored. SPINRec builds on counterfactual evaluation and adapts path integration to sparse recommendation data.

Explanation Methods for Recommenders

  • Model-specific and aspect-based explanations have limited generalization to sparse implicit data, while model-agnostic methods apply broadly but remain insufficiently benchmarked for fidelity.Examples of model-agnostic methods include LIME-RS, FIA, ACCENT, SHAP4Rec, and DeepSHAP.Source: Explanation Methods for Recommenders · p. 2

Fidelity Evaluation in Recommender Systems

  • Recommender explanations have often prioritized persuasiveness or satisfaction, while counterfactual metrics refine evaluation through fixed-length explanations and separate feature roles.Earlier counterfactual approaches could conflate supportive and contradictory features and use coarse fixed-percentage masking.Source: Fidelity Evaluation in Recommender… · p. 2
  • Fidelity evaluation in recommendation remains underexplored, and SPINRec’s extensive empirical study spans both original and refined counterfactual metrics.The paper benchmarks explanation methods across datasets and recommender models.Source: Fidelity Evaluation in Recommender… · p. 2

Path-Integration (PI) for Explainability

  • Path integration can yield more robust explanations by integrating gradients from a baseline to the input, but sparse binary user histories make all-zero baselines unrealistic.SPINRec samples plausible histories from the empirical data distribution to capture presence and absence information.Source: Path-Integration (PI) for Explainability · p. 2

The SPINRec Algorithm

SPINRec adapts path integration to sparse recommendation data by integrating gradients from a baseline to the user profile, then selecting among sampled baselines the explanation with the highest fidelity score. Its data-driven baselines address the limitations of cold-user baselines, while the method’s evaluation cost is dominated by gradient integration and counterfactual scoring.

The SPINRec Algorithm

Setup and Notation

  • Each user is represented by a binary item-interaction vector, and an explanation map assigns each feature a relevance score for the target item's predicted affinity.A feature value of 1 indicates an interaction; higher map values indicate stronger influence on the recommendation.Source: Setup and Notation · p. 2

Challenges in Baseline Selection

Stochastic Baseline Sampling

  • SPINRec samples κ plausible baselines from user histories, computes an explanation map for each, and selects the map maximizing a fidelity metric.The candidate maps can also be averaged, a variant described as generalizing Expected Gradients.Source: Stochastic Baseline Sampling · p. 3

Computational Complexity

  • For each sampled baseline, SPINRec performs J gradient steps and N perturbation-based evaluations; these integration and counterfactual-scoring stages dominate computational cost.The passage notes that J is typically much smaller than N|V|.Source: Computational Complexity · p. 3
  • SPINRec is described as scaling linearly with user features and baseline samples, with parallelizable steps suited to GPU acceleration.The passage contrasts this with SHAP's exponential cost in |V| and LIME's cubic sample complexity.Source: Computational Complexity · p. 3

Counterfactual Fidelity Metrics

The evaluation measures explanation fidelity through counterfactual changes in the target item's ranking or confidence after explanatory features are removed or added. It covers both AUC-based perturbation metrics and fixed-length variants.

Evaluation protocol

  • The evaluation systematically includes original AUC-based metrics and refined fixed-length variants, following the protocols of their respective prior work.AUC variants average model scores as user-vector features are progressively removed or added.Source: Counterfactual Fidelity Metrics · pp. 3–4

Illustrative Example

Removal metrics

  • Removing top-Ke features tests whether the target item drops from top-Kr recommendations, how much its confidence falls, and how its rank degrades.POS@Kr, Ke and CDCG@Ke are lower-is-better metrics; DEL@Ke is also lower-is-better.Source: Counterfactual Fidelity Metrics · p. 4

Insertion metric

Experimental Setup

The experiments compare post-hoc explanation methods across three recommendation architectures and three implicit-feedback datasets. They follow an established evaluation protocol, add Pinterest and additional fidelity metrics, and target each test user's top recommendation.

Experimental Setup

  • The setup extends Barkan et al. (2024) with Pinterest, additional fidelity metrics, and a broader set of explanation baselines, using held-out validation data for hyperparameter tuning.Pinterest results and final hyperparameters are reported as available in the public repository.Source: Experimental Setup · p. 4

Recommendation Models

  • Evaluation spans MF, VAE, and NCF recommendation models.The MF variant derives user embeddings directly from interaction vectors; VAE reconstructs user-item vectors, and NCF combines matrix factorization with multilayer perceptrons.Source: Recommendation Models · p. 4

Datasets

  • ML1M, Yahoo! Music, and Pinterest are binarized as implicit feedback and split by user into 80% training and 20% test sets, with another 10% of users withheld for tuning.Test explanations target each user's top recommendation.Source: Datasets · p. 4

Baselines and Methods

  • The baseline comparison includes heuristic, perturbation-based, learning-based, and influence-function explainers, alongside vanilla path integration and SPINRec.SPINRec combines path integration with fidelity-optimized stochastic baseline sampling; vanilla PI omits that sampling, and the passage reports full SPINRec significantly outperforms the ablated method across settings.Source: Baselines and Methods · p. 4
  • LXR is described as a state-of-the-art fidelity-aware baseline that trains an auxiliary explainer network, while requiring pretraining.The illustrative figure depicts SPINRec-identified history items whose masking sharply lowers “The Lion King” recommendation rank.Source: Baselines and Methods · p. 4, Figure 1 · p. 5

Counterfactual Evaluation Results

SPINRec leads baselines in AUC-based and fixed-length counterfactual fidelity evaluations. Plain path integration is competitive, while stochastic sampling further improves fidelity, especially for VAE and NCF.

Counterfactual Evaluation Results

  • Across all models and datasets, SPINRec significantly surpasses LXR and FIA on AUC-based metrics (p ≤0.01).The significance is reported using a paired t-test.Source: Counterfactual Evaluation Results · p. 5
  • Across fixed-length configurations, SPINRec outperforms all baselines, despite tighter margins and occasional ties at K_e=2 and 3.The tests use K_e values of 2, 3, and 4 and ranking cutoffs K_r of 5, 10, and 20; K_e=1 is omitted for instability and limited discriminative power.Source: Counterfactual Evaluation Results · p. 5

Ablation Study: Plain vs. Stochastic Path Integration

Conclusion

SPINRec applies path integration to recommendation with stochastic baseline sampling tailored to sparse, binary user–item data. Its evaluation finds consistent outperformance of existing approaches across multiple models, datasets, and fidelity metrics.

Conclusion

  • SPINRec is presented as the first model-agnostic explanation method applying path integration to recommender systems.It combines path integration with stochastic baseline sampling tailored to sparse, binary user–item data.Source: Conclusion · p. 6
  • Across multiple models, datasets, and fidelity metrics, SPINRec consistently outperforms existing approaches and establishes a new benchmark for fidelity-aware recommendation explanations.Source: Conclusion · p. 6
  • The authors hope the work encourages further exploration of fidelity-focused explainability in recommendation.Source: Conclusion · p. 6
Loading 2511.18047v1…