Generating Knowledge-based Explanation for Recommendation from Review

Reasonable explanation is helpful to increase the trust and satisfaction of user to the recommender system. Among many previous studies, there is growing concern about generating explanation based on review text.

Collaborative filtering is one of the most successful approaches to predict user's preference. However, most of them suffer from data sparsity problem. Researcher often utilizes auxiliary data to address this problem, such as review, knowledge graph (KG), image and so on. Some researchers have proven that recommendation accuracy can be improved via incorporating rating and review data. Besides, neural network is also applied to learn more powerful representations for user and item from the review data. For example, convolution neural network (CNN) is used to extract representation from review text by using convolutional filters. Recurrent neural network (RNN) is another widely used model, which can encode the sequential behaviours as hidden states. However, most of them lack the ability to generate explanation.

In order to generate explanation, there are two main approaches are used, i.e., template-based approach and generation-based approach. It is usually necessary for the templated-based approach to define serval templates. Then, these templates will be further filled with different personalized features/words. Although they can offer readable explanations, they rely heavily on pre-defined templates. It causes large manual efforts, limiting their explanation expression. Due to the strong generation ability of natural language model, the generation-based approach is capable to generate explanation without templates, which can largely enhance the expression of the generated sentence. Although they can generate more free and flexible explanation, the explanation might tend to be uninformative.

To tackle these challenges of the above-mentioned work, we propose a Generating Knowldge-based Explanation for Recommendation from Review (GKER) to provide informative explanation. Unlike the traditional generation-based approach with a multi-task framework, we design a single-task framework to simultaneously model user's preference and explanation generation. The multi-task training usually needs more manual effort and time overhead. In this unitary framework, we inject the user's sentiment preference into the explanation generation, aiming at capturing the user's interest while producing high-quality explanation. Specifically, we build three graphs, including a bipartite graph, a KG and a co-occur graph. All of them are integrated to form a unitary graph, thus bringing the semantic among user-item interaction, KG and review. Based on this integrated graph, it is possible to learn more effective representations for user and item. To make better use of the integrated KG, a graph convolution network (GCN) is utilized to obtain improved embeddings due to its superior representation learning ability. We argue that these embeddings can contain more semantic interaction signals with the help of the integrated KG and GCN. After obtaining these extensive embeddings, a multilayer perceptron (MLP) layer is further employed to capture non-linear interaction signals between user and item, aiming at predicting user's rating accurately. The predicted rating would be regarded as a sentiment indicator to explore why the user likes or dislikes the target item. To investigate the association between sentiment indicator and the related review data, a transformer-enhanced encoder-decoder architecture is designed to produce informative and topic-relevant explanation. Besides, the aspect semantic is added in this architecture through an attention mechanism. In this framework, the transformer is utilized as a "teacher" model to supervise the generation of the encoder-decoder process. Finally, experiments conducted on three datasets have shown the state-of-the-art performance of GKER.

There are some research issues for discussion: 1) although KG is a useful tool for recommendation accuracy and explainability, it is always incomplete in the real world. Hence, it is worth completing it for the recommendation. 2) Besides, as for explainable, it still needs more metrics to evaluate the quality of its explanation.