| With the development and maturity of emerging Internet technologies,the explosive growth of Internet data has made it more and more difficult to find valuable information from massive amounts of data,which has created the problem of "information overload".The recommendation system is a product that emerged in the era of big data.It can dig out information that users are interested in from a large amount of data and recommend corresponding products according to users’ preferences.With the development of deep learning,current recommendation methods alleviate the data sparseness and cold start problems of traditional recommendation models,but they still face many new challenges.For rating prediction tasks based on review text,most models ignore the interpretability of the prediction results.Due to the black-box characteristics of deep neural networks,it is difficult to convince the prediction results,so accurate prediction results and reasonable explanations become current research hotspots in the field of recommender systems.Existing research models mostly use static and independent methods to extract the potential feature representations of users and item reviews,and represent user preferences as static feature vectors.Only the two entity features interact during the rating prediction stage,ignoring the potential features.The correlation of not only may produce large deviations to the prediction results,but also is not accurate enough for the interpretability of the prediction results,and cannot fully consider the user’s preferences and item characteristics.In response to the above-mentioned problems,this article starts research from the following aspects:(1)This paper proposes an interpretable recommendation method based on interactive attention.Interactive attention is used to study the correlation between user reviews and item reviews,so that users and items can interact dynamically.The hidden vector features extracted by users and items change due to different matching targets.The attention mechanism used by the model generates corresponding weights for the review words.Observing these values,mining words containing important information,and highlighting emotional words or item attribute words from the reviews play an explanatory role in predicting the score.(2)Based on the research of interactive attention,a fine-grain interactive interpretable recommendation method is proposed.The user reviews and item reviews are extracted into multiple aspects,and the degree of matching between the user’s viewpoint and the item characteristics and the importance of the aspect are determined from the aspect level in a fine-grained manner.While improving the accuracy of predictive rating,we can further observe the matching degree of each user’s viewpoint-item aspect pair according to the attention weight of aspect level,and find appropriate reasons to explain the predictive rating result.Compared with the model that only interprets the prediction results from the review level,this model judges the user’s preference for a certain aspect of the item more intuitively and concisely,and provides an interpretable recommendation result from the aspect level. |
