EXPLAINABLE RECOMMENDER WITH GEOMETRIC IN-FORMATION BOTTLENECK

Abstract

Explainable recommender systems have attracted much interest in recent years as they can explain their recommendation decisions, enhancing user trust in the systems. Most explainable recommender systems rely on human-generated rationales or annotated aspect features from user reviews to train models for rational generation or extraction. The rationales produced are often confined to a single review. To avoid the expensive human annotation process and to generate explanations beyond individual reviews, we propose an explainable recommender system trained on reviews by developing a transferable Geometric InformAtioN boTtleneck (GIANT), which leverages the prior knowledge acquired through clustering on a user-item graph built on user-item rating interactions, since graph nodes in the same cluster tend to share common characteristics or preferences. We then feed user reviews and item reviews into a variational network to learn latent topic distributions which are regularised by the distributions of user/item estimated based on their distances to various cluster centroids of the user-item graph. By iteratively refining the instance-level review latent topics with GIANT, our method learns a robust latent space from text for rating prediction and explanation generation. 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 rating prediction accuracy.

1. INTRODUCTION

Typically, a recommender system compares users' preferences with item characteristics (e.g., item descriptions or item-associated reviews) or studies user-item historical interactions (e.g., ratings, purchases or clicking behaviours) in order to identify items that are likely of interest to users. In addition to predictive performance, interpretable recommenders aim to give rationale behind the rating given by a user on an item (Ghazimatin et al., 2020; Zhang et al., 2020) . Most existing interpretable recommenders can either generate rationale or extract text spans from a given useritem review as explanations of model decisions. Both rationale generation and extraction require annotated data for training, e.g., short comments provided by users explaining their behaviours of interacting with items, or annotated sentiment-bearings aspect spans in reviews (Zhang et al., 2014; Ni et al., 2019; Chen et al., 2019; Li et al., 2020; Tan et al., 2021a) . We argue that generating explanations based on a specific user-item review document suffers from the following limitations. First, some reviews may be too general to explain the rating, rendering them useless for explanation generation. For example, the review 'I really like the smartphone, will recommend it to my friends' does not provide any clue why the user likes the smartphone. Second, features directly extracted from a review document may fail to reflect some global properties which can only be identified from implicit user-item interactions. For example, meaningful insights could still be derived from reviews towards items that are not directly purchased/rated by a user but preferred by other like-minded users. Finally, explanation generation model from user/item reviews are often supervised by human-annotated rationales, which are labour-intensive to obtain in practice. To address the aforementioned limitations, we propose an AutoEncoder (AE) framework with variational Geometric InformAtioN boTtleneck (GIANT) to incorporate the prior from user-item interaction graph to refine the induced latent factors of user and item, and generate explanations in an unsupervised manner. For a user-item pair, all reviews written by the user and reviews posted on the 

U-I Graph

Based on the reviews from item: This tape is isn' t transparent, but it is strong. Based on the reviews from like-mined users: Fiberglass threads, impossible to tear by hand, and strong adhesive that will never fall off. Will recommend this item 𝑢 , item are fed into two separate encoders to infer latent factors, and predict rating based on the match of their latent factors. The latent variables are supposed to capture the key semantic information to recover the original review text written by the user on the item. Different reviews are assigned to different latent dimensions according to posterior distributions and we extract the representative words in each dimension to summarize the cluster topic. The explanations are thus from the reviews related to the topic of the assigned cluster (See in §6.3). Our proposed framework is illustrated in Figure 1 . The geometric regularisation refers to the clusterbased distance with Gaussian variance, which is derived by firstly clustering the users/items in the user-item interaction graph, and then calculating the distribution of a user/item as its distance to each cluster centroid by a Gaussian kernel ( § 4.1). To impose the regularisation on the latent variables of the AE framework taking the set of user reviews and item reviews as input, we apply the KL divergence to minimize the discrepancy between the induced posterior distribution of users z (u) , and items z (i) , and the geometric regularisation as the prior ( §4.2), after linking the cluster to the review text encoder via prior-centralisation ( §4.2). Experimental results on the three commonly used benchmarking datasets show that the proposed method achieves performance comparable with several strong baselines in recommendation. Moreover, the quantity and the quality analysis in interpretability show that our method can generate coherent, diverse and faithful explanations.

2. RELATED WORK

We review recommender systems, with particular attention to those built on Variational Autoencoder (VAE) (Kingma & Welling, 2014) or offering explanations of recommendation decisions. Recommender with variational autoencoder Text data such as user reviews, item or brand descriptions could be important for developing a high-quality recommender system as they can be exploited to address the sparsity issue in user-item interactions. HFT (McAuley & Leskovec, 2013) and CTR (Wang & Blei, 2011) adopted Latent Dirichlet Allocation (LDA) (Blei et al., 2003) to extract latent topics from review text. VAE can also be used for text modeling due to its ability in extracting latent and interpretable features (Fei et al., 2021; Truong et al., 2021; Wang et al., 2020) . Truong et al. (2021) argued that the commonly used isotropic Gaussian in VAE is over-simplified and proposed BiVAE by introducing constrained adaptive prior (CAP) for learning user-and itemdependent prior distributions. More recently, review information is used to disentangle the latent user intents at the finer granularity, for example, DisenGCN (Ma et al., 2019a) and DisenHan (Wang et al., 2020) used the graph attention mechanism to differentiate multiple relations and features. Explainable Recommender System One popular interpretable method is to train a language generation model with the ground-truth explanations supplied, which can be the first sentence of a given review or human-annotated text spans in the review text (Li et al., 2017; Chen et al., 2019; Ni et al., 



Don't miss the orange tape • This tape really fits the bill • This tape would be perfect • I don't think you could go wrong with this product.Reviews from item 𝑖 "

Figure1: An encoder-decoder structure with a geometric information bottleneck regularisation, which is derived from the U-I interaction graph and used as a prior imposed on Z. The latent variable z thus can capture the geometric affiliations in graph and be the soft cluster assignment distributions. It enables the use of like-minded users u 3 and similar items i 2 and for generating explanations beyond the input user-item pair (i 1 , u 2 ). Existing rationale extraction methods can only extract indicative words (shown in red) from the given review pair. Our method can generate rationale from reviews written by other like-minded users or on similar items, which are assigned into the same latent dimension/cluster.

