Study On Collaborative Filtering Recommendation Based On Textual Content And Contextual Information

As a main way to solve the information overload problem, recommender systems are proved by related research to be helpful in stimulating sales, accelerating bundle and improving customer satisfaction. Collaborative filtering is the most widely used recommendation technique and has the advantages including simplicity, interpretability,efficiency and stability. However, traditional collaborative filtering depends on the rating data, and because there are several defects for ratings, collaborative filtering faces some challenges, including rating's sparsity, dynamics, and the lack of context information.The challenges of collaborative filtering are especially serious for two types of products (or services): high-involvement products and location-based services. High-involvement products are defined as the durable products with high value. For theseproducts, user's purchase histories are few. Thus, the sparsity problem is more serious.And there is lack of research about high-involvement products' rating dynamics. On the other hand, for the location-based service recommendation, recommender systems should generate the context-aware results; however, the ones based on the ratings cannot meet this requirement. So, in this thesis,we try to introduce two kinds of user-generated content: textual reviews and check-in data, and design the collaborative filtering methods based on them to alleviate the problems caused by only using numeric ratings.The research content and contributions of the thesis include:(1) This thesis conducts empirical studies on high-involvement products' review dynamics from two perspectives of time and order, filling the related research gap. We analyze the review dynamics from the level of numeric ratings and textual reviews. For the rating dynamics, we first give the visual display of rating dynamics to preliminarily determine its existence. Then, we model the relationship between the ratings and two variables: time and order, and prove the existence of temporal and sequential dynamics.At last, we analyze the origin of both dynamics based on the self-selection theory and motivation theory. On the other hand, at the level of textual reviews, we first extract user's feature-level opinions by sentiment analysis techniques. Then, as the rating dynamics, we conduct the analysis from the perspectives of time and order, and give the explanations at last.(2) Focusing on the high-involvement products,we propose the review-based hybrid collaborative filtering methods to solve the rating sparsity and dynamics problem of high-involvement products. At first, to solve the sparsity problem, we mine the topics, features and the corresponding opinions to infer the virtual ratings. Different from previous research which mainly uses the overall sentiment of reviews, in the thesis,we leverage the feature-level sentiment and construct the Item-Topic rating matrix. In addition, when designing the recommendation method, we consider the dynamics of ratings and user opinions, and give the solutions. At last, the experiment results on the high-involvement products show that the proposed method can effectively improve the recommendation performance.(3) Considering the temporal and spatial effects of mobile user's check-in behavior, we design the context-aware location-based services recommendation model.In particular, we improve the computation of location similarities from temporal and spatial aspects. In the temporal aspect, we introduce the time factors into the User-Location check-in matrix by time partition. Then, considering the non-uniformness and consecutiveness properties of the check-in time,we compute the similarities between different time intervals to alleviate the sparsity problem because of time partition. In the spatial aspect, we propose the concept of spatial proximity. We first find each user's active regions according his/her check-in histories. Then, considering the active regions and the check-in ratios in active regions, we give the definition of the spatial proximity.The experiment results prove that, compared to the baseline methods, the proposed location recommendation method in the thesis achieves better performance.The implications of this thesis is significant. With regards to the theories,this work provides a deeper understanding the properties and variation patterns of two types of user-generated content: online reviews and check-in data, and gives an efficient way to solve the challenges of collaborative filtering. With regards to the practice,two types of recommendation frameworks designed in this thesis can provide references for related e-commerce enterprises.