Explainable Recommender with Geometric Information Bottleneck

TL;DR

GIANT introduces a geometric information bottleneck framework for explainable recommendations by coupling a text-based variational model with a geometry-derived prior from the user-item interaction graph. The geometric prior is obtained via clustering of graph embeddings using a Gaussian kernel, yielding cluster-probability priors that regularize the latent factors learned from reviews through a transferable information bottleneck objective $O_{IB} = I(\bar{X};Z) - \beta \cdot I(X;Z)$. This approach enables sentence-level explanations by selecting reviews from like-minded clusters and produces semantically coherent latent topics, while maintaining competitive rating prediction and improved interpretability across BeerAdvocate, Digital Music, and Office Products datasets. The method demonstrates that global community structure can be incorporated into text-based encoders to enhance interpretability without sacrificing predictive performance, offering practical benefits for transparent and user-trustworthy recommender systems.

Abstract

Explainable recommender systems can explain their recommendation decisions, enhancing user trust in the systems. Most explainable recommender systems either rely on human-annotated rationales to train models for explanation generation or leverage the attention mechanism to extract important text spans from reviews as explanations. The extracted rationales are often confined to an individual review and may fail to identify the implicit features beyond the review text. To avoid the expensive human annotation process and to generate explanations beyond individual reviews, we propose to incorporate a geometric prior learnt from user-item interactions into a variational network which infers latent factors from user-item reviews. The latent factors from an individual user-item pair can be used for both recommendation and explanation generation, which naturally inherit the global characteristics encoded in the prior knowledge. Experimental results on three e-commerce datasets show that our model significantly improves the interpretability of a variational recommender using the Wasserstein distance while achieving performance comparable to existing content-based recommender systems in terms of recommendation behaviours.

Figures (3)

• Figure 1: The proposed geometric prior reflects the U-I distances in their interaction graph, frequently interacted users/items are assigned into the same cluster (illustrated in node colour). Existing extractive methods can only identify indicative words (shown in red) from the given review pair $(x_i,x_u)$, while our method can extract sentence-level reviews written by other like-minded users or on similar items, which are assigned into the same latent dimension/cluster, i.e., we extract the sentences from the review by user $u_3$ which is in the same cluster as user $u_2$, as the reviews of $u_2$ are too general to summarise the key features.
• Figure 2: Top: The Comprehensiveness values by removing the top $k$ most important latent dimensions identified, $k \in \{3, 5, 10, 15, 20,30\}$. Ous model uses the left y-axis, the two baselines use the right axis. Bottom Table: Relative performance changes after subtracting the changes caused by random removal.
• Figure 3: Extracted interpretations from GIANT (upper) and WassersteinVAE (bottom) for the item, a tape. In the example generated by GIANT, both the reviews/rationales from the user and item side focus on the 'sturdy and strong', and reveal lots of important aspects for the recommendation. While the user reviews extracted by WassersteinVAE, e.g., 'go wrong' fail to capture the key factors, and the extracted item reviews are even contradictory, i.e., 'won't be using' to the recommended behaviour.