| Summary: | 碩士 === 國立成功大學 === 資訊管理研究所 === 104 === The excessive amount of user-generated reviews results in the difficulty of extracting the relevant information. A recommender system is a solution to help users get the accurate information efficiently. Most of the existing recommender systems neglected emotions expressed by users in the reviews. It caused the failure of predicting user preferences accurately. On the other hand, the cold start problem and the model scalability are the two thorny problems to the recommender system. Cold start exists while lacking initial ratings; and model scalability is the capability of a model to cope with the high-dimensional data. These problems may mislead the recommendation, and users are not satisfied with the results accordingly.
A personalized recommender system is proposed to mitigate the negative effects the aforementioned problems cause. After extracting user preference, the social influence network is built accordingly. The predicted ratings are estimated based on the importance of users. Also, ontologies are applied to integrate the extracted features into topics for the sake of dimensionality reduction; and topics mentioned in the reviews are displayed as a form of topic map.
In addition, this thesis designed a number of experiments to validate the effectiveness of the proposed method. RMSEs and MAEs of all relationships are close 1, and most of the F1 measure is larger than .8. Both of them indicate that the proposed method is able to estimate the unknown ratings well with the help of social influence. Among all, the combinations of relationships perform better.
|
|---|
