一個新的資料特徵產生方法應用於手機影片推薦之使用者分群之研究

Abstract

在現行的推薦系統來說，Amazon已經證實了透過社群推薦的成功，但是在手機影片推薦系統當中，由於手機影片更新的速度快而且手機的螢幕小的原因使得新的手機影片無法有大量曝光的機會。根據這樣的限制，內容導向過濾偕同合作式過濾推薦方法(Content-based collaborative filtering)就被應用來解決此問題。而原本用來描述影片內容的標籤並不適合用來描述使用者的特徵，這些標籤之間存在著相依或重複性等問題而導致不平衡的使用者分群結果，因此，如何去精煉大眾分類法的標籤成為獨立發散的屬性以用來進行使用者特徵的分群分析就變成很重要的一個議題，在這篇論文當中，一個用於使用者分群的創新的屬性產生方法會透過篩選掉多餘無用的標籤、收縮支配性高的標籤以及歸納隱含性的屬性等方式來凸顯使用者的特徵。而在實驗當中，會透過真實10906位手機用戶的28249筆交易資料，透過training data與testing data來驗證。且實驗的結果將證明我們的研究能夠得到更佳的使用者分群，並且能夠提昇使用者族群之間的差異性，以及提昇推薦的命中率。

In recent years, the growth of Amazon proved the success of social recommendation. However, in the mobile video recommendation system, the nature of frequent updates of entertainment video and the small screen doesn’t allow new contents to have many opportunities for exposure. To solve the issue, the Content-based Collaborative Filtering (CBCF) recommendation approach is applied to solve new item problem. CBCF relies on attributes to characterize users’ preferences and makes recommendations according to the log of clustered users with similar interests and the descriptions of contents for users’ preferences. Since the common folksonomy-based tag system for video are used to describe the properties of video content but not users’ characteristics, the tag dependency problem causes poor user clustering result. Therefore, how to refine the folksonomy-based tags to independently and identically distributed attributes for user characteristics clustering analysis becomes an important issue. In this thesis, a novel attribute generation method based on a taxonomy-based attribute system for user clustering is proposed to reveal the users’ characteristics which are screening the redundant tags, shrinking the dominated attributes and generalizing the implicit attributes. In the experiment, the 28249 transactions with 10906 users of the real mobile phone customers have been adopted as training and testing data. The experimental result shows our approach can obtain the representative user characteristics in the clustering results to improve the recommendation.