Research On Multi-granularity Co-Attention Recommendation Combined With Auxiliary Comments

Digital construction in China has benefited a lot from the rapid growth and widespread application of big data technology.However,numerous data generated during the process and inevitably introduced the serious ”information overload” problem.To help users choosing items that match their preferences from mass data,the Recommendation System emerges.As the carrier of problems,suggestions and attitudes of users,reviews are widely concerned for their capacity of excavating users' preferences effectively.But existing review-based recommenda-tion methods didn't consider the sparsity of data,Also,the outputs can be hardly explained.To solve these problems,we proposedthe Multi-Granularity Co-Attention Recommendation Model Combined with Auxiliary Comments(MCRA).Our work is concluded as follows:(1)To improve the interpretability of recommendation results,MCRA constructed net-works for both review-level and aspect-level.At the review level,representation of user reviews are generated via CNN,then attention mechanism is introduced to evaluate the importance of each review with the combination of review and item ID.After that,weight related to each re-view will be learned with remote supervising.In the aspect level,CNN and gate mechanism are used to learn the representation of user reviews under different aspect categories.The rep-resentations along with the user review weights generated from the attention mechanism under review-level finally produce a weighted sum,which is the final aspect representation of users.The review weights of the item and the representation of the item aspect follow similar rules.Considering the dynamic aspect association for user-item pair,our model used the Co-Attention mechanism to evaluate the importance of user-item pair in each aspect.Finally,the rating pre-diction will be generated through LFM with aspect representation of user-item pairs and the aspect importance estimations.(2)As for the sparsity of comment data,MCRAintroduced auxiliary reviews into the aspect level.Auxiliary reviews are selected from similar user reviews under the guidance of outputs from the attention mechanism in the review level.Users and aspect representation of auxiliary comments will be combined to generate the final aspect representation of each review.We designed and conducted widely contrast experiments in Amazon and Yelp with real datasets from four different fields to evaluate the performance of our model.Also,we used fusion experiments to verify the effectivity and necessity of the aggregation for our model and conducted hyper-parameter experiments to verify the impact of hyper-parameter on the performance of our model.Finally,examples are used to explain the interpretability of our model's outputs.