基於矩陣分解與主題模型之隱含回饋文章推薦

Abstract

推薦系統被廣泛的運用以解決資訊過載的問題,且大部分的推薦系統主要針對使用 者有評分的顯性回饋資料進行推薦。然而現實生活中所能收集到的資料記錄大部分為隱 含回饋,像是購買記錄、瀏覽紀錄、點擊記錄...等。而針對這樣的資料,推薦系統很難 判別使用者的喜好,因為對於與使用者沒有相關記錄的項目,無法判別使用者對於該項 目是沒有興趣或沒有注意到。大部分使用隱含回饋資料的推薦系統都將未記錄的資訊視 為負面回饋,然而,這樣的假設會誤把使用者喜歡的項目視為負面回饋,使得推薦的結 果產生偏差。 為了改善上述的問題,本研究提出結合主題模型與矩陣分解的文章推薦方法,考慮 文章中的主題機率分佈以及相似使用者的喜好,試圖找出使用者可能會給予負面或正面 回饋的文章,並且解決資料過於稀疏的問題,進而利用矩陣分解來預測使用者對於文章 的喜好。實驗結果顯示本研究所提出的方法能夠有效地針對使用者的喜好推薦其感興趣 的文章,並改善過去以隱含回饋為主之推薦系統的準確度。

Recommender systems have been applied in many domains to solve the information- overload problem, and most of them make recommendations based on explicit data which expressed ratings in different scores. However, there are a lot of implicit data in the real world, such as users’ purchase history, click history, browsing activity and so on, and it is difficult to find users’ preferences based on this kind of data. To solve this problem, most researches set all missing data as negative assuming each user only likes few items. Nevertheless, recommendation methods based on such assumption may mistakenly assign the wrong negative examples and lower the recommendation quality. In this work, we proposed a novel recommendation method, which incorporates topic model and matrix factorization. The content of documents and similar users’ preferences are used to predict the negative examples. Moreover, since the collected data is very sparse, the potential positive examples are also predicted according to similar users’ preferences. Finally, matrix factorization is adopted to make predictions on items. The experimental results show that the proposed approach achieves better performance than other recommender systems with implicit feedback.