SoREX: Towards Self-Explainable Social Recommendation with Relevant Ego-Path Extraction

TL;DR

SoREX tackles the explainability gap in graph-based social recommendation by introducing a self-explainable, two-tower GNN framework that separately models social influence and user-item interactions. It introduces an ego-path based explanation mechanism that converts a target user’s ego-net into dense, candidate-aware, factor-specific subgraphs, enabling comparative explanations across ranked items and tying explanations to predictions through explanation re-aggregation. An auxiliary friend-recommendation task strengthens social embeddings, and multi-task training jointly optimizes both recommendation and social signals. Across four benchmarks, SoREX achieves superior predictive accuracy and provides quantitative and qualitative evidence of interpretable, comparative explanations, supported by extensive ablation and fidelity analyses.

Abstract

Social recommendation has been proven effective in addressing data sparsity in user-item interaction modeling by leveraging social networks. The recent integration of Graph Neural Networks (GNNs) has further enhanced prediction accuracy in contemporary social recommendation algorithms. However, many GNN-based approaches in social recommendation lack the ability to furnish meaningful explanations for their predictions. In this study, we confront this challenge by introducing SoREX, a self-explanatory GNN-based social recommendation framework. SoREX adopts a two-tower framework enhanced by friend recommendation, independently modeling social relations and user-item interactions, while jointly optimizing an auxiliary task to reinforce social signals. To offer explanations, we propose a novel ego-path extraction approach. This method involves transforming the ego-net of a target user into a collection of multi-hop ego-paths, from which we extract factor-specific and candidate-aware ego-path subsets as explanations. This process facilitates the summarization of detailed comparative explanations among different candidate items through intricate substructure analysis. Furthermore, we conduct explanation re-aggregation to explicitly correlate explanations with downstream predictions, imbuing our framework with inherent self-explainability. Comprehensive experiments conducted on four widely adopted benchmark datasets validate the effectiveness of SoREX in predictive accuracy. Additionally, qualitative and quantitative analyses confirm the efficacy of the extracted explanations in SoREX. Our code and data are available at https://github.com/antman9914/SoREX.

Figures (10)

• Figure 1: A toy example illustrating a substructure-based comparative explanation for recommending item $v_4$ to user $u_3$. We enumerate all meta-paths of length no longer than 3 from $u_3$ to items $v_{1}$, $v_{2}$, and $v_{4}$. While $v_{1}$ and $v_{4}$ each have two paths and $v_{2}$ only one, $v_{4}$ is more relevant due to a shorter, socially meaningful path via direct friend $u_{4}$. This highlights how structural signals alone can yield accurate and interpretable recommendations.
• Figure 2: Example of ego-net and ego-path. The 2-hop ego-net of user $u_3$ is highlighted on the left, while the ego-net is transformed into a set of 2-hop ego-paths on the right. Given the length of the ego-path passing by $v_6$ is less than 2, to meet the required length, it is padded with an empty node.
• Figure 3: Overview of our proposed SoREX framework for self-explainable GNN-based social recommendation.
• Figure 4: Hyperparameter sensitivity analysis. The figure presents the effects of varying the key hyperparameters on the model's performance, including the choice of the aggregation function, the multi-task balancing parameter $\gamma$, and the number of layers $k_1$ and $k_2$ in both graphs. The rightmost bar (Jaccard) is significantly better than Cosine and Pearson at the 5% level (paired t-test, $p<0.05$).
• Figure 5: Stability analysis of SoREX with respect to the number of ego-paths on LastFM and Yelp. (a) and (b) report the average and standard deviation of NDCG@10 on LastFM as the number of ego-paths varies. (c) and (d) show the corresponding results for Yelp.