- 學位論文
可改善推薦系統評價預測之使用者分群方法研究
Improving the Prediction Accuracy of Recommender Systems by Clustering Users Based on User and Group Similarity
摘要
推薦系統已經成功運用在許多地方，而過去學者們也不斷提出新方法或架構，希望能得到更好的推薦準確度。而系統在進行推薦時，群組中若含有大量使用者，會導致系統在進行推薦時耗費大量時間成本。為了解決這個問題，本研究提出一套使用者分群演算法，利用目標使用者與群組的平均相似度做為判斷依據，當相似度大於門檻值，則可判斷使用者與群組為相似。其中門檻值的設計我們是採用動態門檻值，透過目標群組的人數與系統中最大群組的人數為基礎，產生動態門檻值，運用此門檻值可以解決固定門檻值所發生的問題，且可以使得系統更有彈性，不論系統中使用者如何增加，這套門檻值也能動態的更改。經由本研究所提出的分群演算法，可以將使用者分成多個群組，群組中的使用者可視為相似使用者，之後系統在進行推薦時，若需要找尋相似使用者，只需要與群組中的使用者做比對，如此一來可以減少系統在進行推薦時的時間，同時也可以增加預測的準確度。我們最後使用MovieLens電影資料庫做為實驗資料，證明此分群演算法比協同式過濾有更好的效率以及準確度。
並列摘要
Recommender systems have become a popular issue in the mid-1990s. Scholars have proposed a lot of method and framework to improve the prediction accuracy. This thesis presents a clustering algorithm according to user and group similarity. We calculate the similarity between users by Pearson correlation coefficient based on user’s ratings. By using this similarity, we can calculate the average similarity between the target user and existing groups. If the average similarity is higher than the threshold which we set, the target user will be clustered on the group that has the highest average similarity. The threshold we propose is a dynamic value based on the size of group. The larger the group, the lower the threshold becomes. The design of threshold could avoid some problem of static threshold and makes our method more flexible. According to the clustering algorithm that we present, it is not only able to improve the prediction accuracy but also decreases the execution time. We present experimental results with the MovieLens data set to show that the proposed method has better performance and prediction accuracy than a pure collaborative filtering.