| Summary: | 碩士 === 國立中正大學 === 電機工程研究所 === 107 === The purpose of this study is to solve the distortion from several causes. The causes are acquiescence bias and central tendency bias, which have been a key issue in psychometrics. Acquiescence bias is sometimes referred to as "yea-saying" which is a kind of response style. This study has developed a new preprocessing method, which changes the rating values from the users who show the acquiescence bias or the central tendency bias. The study proposes the preprocessing method with several machine learning algorithms to boost the performance of a recommendation system. The study evaluates the recommended list with several evaluations, such as Area under the Curve (AUC), precision, recall, Average Precision (AP), normalized Discounted Cumulative Gain (nDCG), novelty and coverage. The evaluation results have shown that novelty and coverage have improved in most algorithms. A higher novelty means that user on this system would be more willing to explore new interests. A higher coverage means less influence caused by cold start problem. The study explains the influence on specific algorithms, Latent Dirichlet Allocation (LDA) and Bayesian User Community Model (BUCM), which are most improved on novelty and coverage. At the end, the study suggests that when the dataset used is a questionnaire, especially in Likert Scale, preprocessing the rating values caused by the users with the acquiescence bias and the central tendency bias would get better results on novelty and coverage.
|
|---|
