추천 시스템 설명 충실도 향상을 위한 확률 경로 통합 기법 SPINRec

📝 Abstract

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 .

💡 Analysis

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 .

📄 Content

Recent advances in recommender systems over the past decade (He et al. 2017;Kang and McAuley 2018;He et al. 2020; Barkan et al. 2019;Barkan, Katz, and Koenigstein 2020;Barkan et al. 2021;Katz et al. 2022) have increasingly shaped personalized decisions across e-commerce, social media, and streaming platforms, making transparency and trust more essential than ever. (Fan et al. 2022). Explainability in these systems is critical not only for user satisfaction but also for accountability, compliance with regulations, and user control. However, while explainable recommendation research is rapidly expanding (Zhang, Chen et al. 2020;Varasteh et al. 2024), most existing work focuses on user-centric aspects such as persuasiveness, clarity, or satisfaction (Kunkel et al. 2019;Tintarev 2025). A critical yet underexplored dimension is fidelity, which measures how accurately explanations reflect a recommender’s actual decision process. Without fidelity, explanations may appear plausible while failing to reveal the true reasoning behind recommendations (Koenigstein 2025).

We introduce SPINRec (Stochastic Path Integration for Neural Recommender Explanations), the first adaptation of path-integration (PI) (Sundararajan, Taly, and Yan 2017) to recommender systems. Unlike prior applications of PI in vision (Kapishnikov et al. 2021;Barkan et al. 2023a,c;Elisha, Barkan, and Koenigstein 2024;Barkan et al. 2025Barkan et al. , 2023d,b) ,b) or NLP (Sikdar, Bhattacharya, and Heese 2021;Enguehard 2023), recommender data is characterized by extreme sparsity and binary-valued interactions, where the absence of an interaction may be ambiguous. Standard PI methods, which integrate gradients from an all-zero baseline, fail in this setting due to weak or misleading attribution signals. Crucially, modern recommenders leverage both observed and unobserved interactions as informative signals. SPINRec addresses this by stochastically sampling plausible user baselines from the empirical data distribution and selecting the explanation that maximizes fidelity. This adaptation enables more stable and faithful explanations tailored to the structure of recommender systems.

To evaluate SPINRec, we conducted extensive fidelity evaluation spanning three model architectures (MF, VAE, NCF), multiple benchmark datasets (ML1M, Yahoo! Music, Pinterest), and a suite of counterfactual fidelity metrics (Barkan et al. 2024;Gurevitch et al. 2025;Baklanov et al. 2025). Our results establish SPINRec as the new state-of-the-art benchmark in recommender systems explainability, with ablation studies confirming the distinct contributions of both pathintegration and our stochastic baseline sampling strategy.

• Introduce SPINRec, the first adaptation of path-integration methods to recommender systems. • Develop a novel stochastic baseline sampling strategy tailored to sparse, binary recommendation data. • Conduct comprehensive fidelity-focused evaluation across multiple architectures and datasets. • Establish SPINRec as the new state-of-the-art for fidelityaware recommendation explanations. As fidelity remains an underexplored yet critical dimension in explainable recommendation (Baklanov et al. 2025;Mohammadi et al. 2025;Koenigstein 2025), we expect this work to lay essential groundwork for future research on trustworthy, model-faithful explanations.

The rapid growth of recommender systems has driven increasing interest in Explainable AI (XAI) methods to ensure transparency, build trust, and enhance user engagement (Tintarev and Masthoff 2022;Zhang, Chen et al. 2020). While a broad range of explanations for recommenders exist, standardized benchmarks for explanation fidelity remain significantly underexplored (Baklanov et al. 2025;Mohammadi et al. 2025;Koenigstein 2025).

Early works typically proposed model-specific explanations, such as those for matrix factorization (Abdollahi andNasraoui 2016, 2017) or inherently interpretable recommender architectures (Barkan et al. 2020Melchiorre et al. 2022;Gaiger et al. 2023;Sugahara and Okamoto 2024). Aspect-based methods attribute recommendations to humaninterpretable item features (e.g., price or color) (Vig, Sen, and Riedl 2009;Zhang et al. 2014;Wang et al. 2018;Li, Zhang, and Chen 2021). However, these methods rely heavily on structured feature availability and are difficult to generalize to implicit or sparse data scenarios.

Model-agnostic explanation methods offer broader applicability by explaining arbitrary recommenders independently of their internal mechanisms. Prominent examples include LIME-RS (Nóbrega and Marinho 2019), influence-based methods such as FIA and ACCENT (Cheng et al. 2019;Tran, Ghazimatin, and Saha Roy 2021), and Shapley-value-based methods (SHAP4Rec and DeepSHAP) (Zhong and Negre 2022;Lundberg and Lee 2017a). While model-agnostic methods enhance generality, their fidelity remains less scrutinized and insufficiently benchmarked.

Explanation fidelity, the degree to which an explanation reflects

This content is AI-processed based on ArXiv data.