A Text-based Explainable Recommender System

With the advent of the Internet,society has entered the era of information overload,and the recommender system has become an important part of online services such as e-commerce websites,social media and news portals.Recommender system aims to ef-fectively mine user interests and address information overload.Existing mainstream rec-ommender systems tend to dig out the interests of users and perform recommendations from the historical interaction between users and items,called interaction-based recom-mender system.Although the interaction-based recommender systems have achieved good performance,they are likely to suffer serious performance degradation when the interaction data is sparse.In addition,the interaction-based recommender systems cannot provide explainable recommendationsIn recent years,the text-based recommender systems have become prevalent to address the issues faced by interaction-based recommender systems.The reviews written by users are one of the textual data sources that reflects the purchasing experiences and sentiment of users(e.g.why the user wants to buy the product and whether the user is satisfied with the product)of users' online consumption in addition to the interaction record.The text-based recommender systems attempt to tackle the sparseness problem of the interaction data by using the review data or item description as new sources of data.And meanwhile,the text-based recommender systems equip the recommendation with explainability based on the extracted semantic informationThis paper focuses on the issues of existing text-based recommender systems,and studies the following three problems(1)Context-aware dynamic user-item representation learningMost text-based recommender systems rely on the parallel architecture to in-fer the latent representation of users and items based on user documents and item documents,respectively.Although this architecture is good at mining the general preferences of users and the basic attributes of items,it is incapable of capturing the shopping context of users,makes the text-based recommender systems fail to capture the fine-grained and complex relationship between users and items.This paper proposes a user-item joint representation learning model with context-aware ability by analyzing the user's shopping context on the e-commerce platform su-percharged with the deep learning technique and the attention mechanism.The key idea is utilizing the semantic interactions between user documents and item documents to learn the context-aware joint representation of users and items,and thus improve the recommendation performance and provide fine-grained semantic explainable recommendation results.(2)Reasoning users' purchased behaviors based on the relationships between user pref-erences and item characteristicsInferring the reason and sentiment behind users'shopping behavior based on users' reviews is one of the major focuses of modern mainstream e-commerce plat-form.The existing explainable recommender systems cannot uncover whether a user likes or dislikes an aspect of an item and to what extent.Therefore,they fail to reason the underlying factors which drive users'shopping behavior.In this paper' we employ the famous capsule network to model the relationships between user preferences and item attributes with respect to the two polarities of sentiment and derives the sentiment and degree carried by users towards different attributes of the same item.Finally,we establish a capsule network-based architecture to explain what you like and dislike for item recommendation.(3)Addressing the sparsity of review dataAlthough text-based recommender systems have alleviated the sparsity prob lem of interaction data,they are incapable of maintaining a robust performance when the text data is sparse.This paper aims to regard the users with similar rat-ing behaviors as the like-minded users and then alleviate the sparsity of text data by extracting the semantic information contained in the auxiliary reviews written by the like-minded users.The goal is to leverage the auxiliary reviews written by the like-minded users as additional textual data to ensure the robustness of the text-based recommender system when the text data is sparse,and meanwhile,tries to utilize the extracted semantic information to improve the recommendation performance.