Explainable Recommendation with Personalized Review Retrieval and Aspect Learning

“Recent years have witnessed a growing interest in the development of explainable recommendation models [1, 2]. In general, there are three different kinds of frameworks for explainable recommendation models, which are post-hoc [3], embedded [4], and multi-task learning methods[5]. Post-hoc methods generate explanations for a pre-trained model after the fact, leading to limited diversity in explanations. Embedded methods, on the other hand, demonstrate efficacy in acquiring general features from samples and mapping data to a high-dimensional vector space. However, since embedded methods rely on historical interactions or features to learn representations, they may struggle to provide accurate recommendations for users or items with insufficient data.

In addition to the two frameworks mentioned above, there has been a utilization of multi-task learning frameworks in explainable recommendation systems, where the latent representation shared between user and item embeddings is employed [1, 5]. These frameworks often employ the Transformer [6, 7], a powerful text encoder and decoder structure widely used for textual processing tasks. While efficient for prediction tasks, they encounter challenges in generation tasks due to limited review content, leading to a significant decline in performance. Furthermore, these previous transformer-based frameworks do not incorporate personalized information and treat heterogeneous textual data indiscriminately. To address these issues, we make adaptations to the existing multi-task learning framework by incorporating two main components: retrieval enhancement, which alleviates the problem of data scarcity, and aspect enhancement, which facilitates the generation of specific and relevant explanations.”

Real-world datasets usually contain redundant reviews generated by similar users, making the selected reviews uninformative and meaningless, which is illustrated in Figure 1. To address this issue, a model-agnostic retrieval enhancement method has been employed to identify and select the most relevant reviews. Retrieval is typically implemented using established techniques, such as TF-IDF (Term Frequency-Inverse Document Frequency) or BM25 (Best Match 25) [8], which efficiently match keywords with an inverted index and represent the question and context using high-dimensional sparse vectors. This approach facilitates the generation of sufficient specific text, thereby attaining enhanced textual quality for the user. Generally, Wikipedia is utilized as a retrieval corpus for the purpose of aiding statement verification [9, 10]. Here, we adopt a novel approach wherein the training set of each dataset is utilized as the retrieval corpus. By integrating this component into our framework, we are able to generate sentences with more specific and relevant details. Consequently, this enhancement facilitates the generation of explanations that are more accurate, comprehensive, and informative at a finer granularity.

Moreover, users rarely share a common preference [11]. Therefore, aspects [12], extracted from corresponding reviews, can be utilized to assist in the modeling of user representation. The incorporation of aspect enhancement has resulted in not only improved prediction accuracy but also more personalized and user-specific text during the text generation process.

The main contributions of our framework are as follows:

• In response to the problem of insufficient historical reviews for users and items in explainable recommendation systems, we propose a retrieval enhancement technique to supplement the available information with knowledge bases obtained from a corpus. To the best of our knowledge, this study represents the first application of retrieval-enhanced techniques to review-based explainable recommendations.

• We propose a novel approach wherein different aspects are selected for individual users when interacting with different items and are subsequently utilized to facilitate the modeling of user representation, thereby leading to the generation of more personalized explanations.

Users’ preferences are often reflected in their reviews. To better represent users, we need to select the most important aspects of their reviews. Specifically, we first extract aspects from each user and item review using extraction tools. The extracted aspects from user reviews represent the style of the users in their reviews, while the extracted aspects from item reviews represent the most important features of the item. We aim to identify the most important aspects that users are concerned about in their reviews. It is worth noting that users’ interests may vary in different situations. For example, when choosing a hotel, a user may care more about the environment. Whereas, price is a key factor to consider when buying a mobile phone.

6 CONCLUSION

In this paper, we propose a novel model, called ERRA, that integrates personalized aspect selection and retrieval enhancement for prediction and explanation tasks. To address the issue of incorrect embedding induced by data sparsity, we incorporate personalized aspect information and rich review knowledge corpus into our model. Experimental results demonstrate that our approach is highly effective compared with state-of-the-art baselines on both the accuracy of recommendations and the quality of corresponding explanations.