Research On Automobile Recommender System Based On Sentiment Analysis Of Comments And Probabilistic Model

With the implementation of the big data strategy, Internet applications, led by E-commerce platform, are now deeply ingrained in modern lives, and also transform the development mode of traditional industries including the automobile. The development of Internet platform and information technology offers consumers more conveniences and reference information for purchasing automobiles. But in the meantime, the problem of information overloads is coming from bad to worse.Recommender system is a kind of active extraneous information filtering technology. With the effectiveness and efficiency in ordinary item recommendations, the improved algorithms of content-based and collaborative filtering are not only widely applied in the major Internet platform, such as e-commerce, Internet finance and social network, but also become one of the focuses in academic research. However, for the personalized recommendation in some domain-specific scenarios, such as house, automobile and financial product, the effectiveness of collaborative filtering and content-based algorithms is always hard to meet user’s requirements.Based on the deep analysis of problems existed in the automobile personalized recommendation tasks of traditional recommender systems, this paper proposes APRS which is a novel automobile personalized recommender system using a probabilistic model. With this recommender system, we can assist consumers with less knowledge about automobiles to conduct purchasing decision-making. The main contents of this thesis are as follows:(1) According to the features of Chinese products on-line reviews, we build the domain dictionaries and discover the part-of-speech(POS) patterns of opinion clauses firstly, and then propose the SAOSP which is an opinion mining and sentiment analysis algorithm based on product features. In order to complete each work of opinion mining and sentiment analysis effectively, our algorithm analyzes the product reviews at the granularity of the clause. As a result, we supplement the product information based on objective data facts, and lay the data foundation for the recommendation model.(2) After analysis of consumers’ behaviors in the purchasing process of domain-specific product, this paper proposes a probabilistic model to describe the purchasing process of the automobile. We introduce user individual preference, user community from Internet and social community in reality into our model, as these factors can affect the consumers’ purchasing decision-making. And then the parameters in the model are estimated by matrix projection.(3) Given the high complexity of domain-specific products and the lack of related domain knowledge among ordinary consumers, the recommender system we proposed guides the consumers’ requirements in a session explicitly and bridges the gap between the requirements of consumers and the features of products by our probabilistic model. We also provide the ranking formula for recommending the Top-K products to a user.(4) By introducing the improved algorithms, FMM and TR, based on collaborative filtering for comparison, we conduct a series of experiments with real-life data sets for the proposed recommender system. Our performance study confirms the effectiveness and accuracy of our recommender system, APRS, for automobiles.