Recommendation With Structure And Text Information Via Heterogeneous Information Network

Recommender system has been well investigated in the past years.As a typical representative,Collaborative Filter(CF)has been widely used to recommend items for customers based on their historical purchase behaviors or ratings on items.HoweverCF-like models often give recom-mendation with low accuracy when the interaction information between users and items are sparse.Latent factor models have been widely used for recommenda,tion.Most of latent factor models mainly utilize the rating information be-tween users and items,although recently extended models may use some auxiliary information(e.g.,social network,heterogeneous information).Essentially,previous latent factor models for recommendation generally lean a unified latent factor through rating interaction between users and items to represent each user or item without distinguish different aspects of the users' preference.However,in rich social media situations,latent features of users and items may stem from different aspects,e.g.,latent features of users and movies from the genres or directors of movies in movie recommendation.In this paper,we first study the aspect-level latent factor for recommendation,and propose the Aspect-level Collaborative Filtering(ACF)model.Through modelling rich objects and semantics in recommendation system as a heterogeneous information network,the ACF model can extract different aspects hrough meta paths.Further-more,a deep neural network is designed to learn different aspect-level latent factors and integrate these latent factors for the recommendation.Extensive experiments on three real datasets show that the proposed ACF model can significantly improve recommendation performances,compared to traditional latent factor models and recent neural network models.In order to handle the cold-start problem,we can leverage lots of addit,ional available data information that can help to alleviate the issue of insufficient interaction information between users and items,e.g.,the review text.In recommender system,besides ratings on items,customers often provide review text to explain the reasons about their ratings about the items.In this paper we develop a novel Representation Learning with Depth and Breadth(RLDB)model for better recommendation by ef-fectively exploiting the user-item interaction information and the users'review texts on items.Specificallywe design a heterogeneous network embedding method and convolutional neural network based method to learn feature representations of users and items from user-item interac-tion structure and review texts,respectively.Furthermore,an end-to-end breadth learning model is proposed through employing multi-view ma-chine technique to learn features and fuse these diverse types of features in a uniform framework.To verify the effectiveness of our proposed rep-resentation learning model,we conduct extensive experiments using three real-world datasets and compare with the other typical baselines.The empirical study clearly demonstrates that our model outperforms all the other methods in these datasets.