以主題模式建構使用者喜好特徵檔之個人化新聞推薦

Abstract

個人化推薦系統在幫助使用者找到合適且有用的標的物上，是非常重要的輔助工具。然而因著情況的改變，讀者的閱讀喜好也可能會隨著時間產生變化，因此要在線上新聞網站上做文章的推薦，就必須要掌握使用者的動態喜好。並且，大多數現存的新聞推薦方法僅單一擷取讀者閱讀的文章相關資訊，例如：內文、分類、關鍵字等作為分析的資料集，本研究所提之方法不僅做讀者對文章的直接喜好分析，也另將網站中的抽獎參與資訊納入計算，藉以間接發掘使用者的潛在興趣。 本研究提出一個人化新聞推薦方法，資料集來源「妞新聞」網站：NIUS news (www.niusnews.com)，是一個以女性為導向的中文線上新聞網站。為要提供準確的推薦，我們採用主題模式技術之一的Latent Dirichlet Allocation (LDA)方法來對新聞內容以及抽獎物之敘述進行主題特徵的分析，所產生之主題機率向量形成了文件及商品特徵檔，再觀察使用者與文章及商品的互動情形，建立出使用者特徵檔案。最後再藉由相似度的比較建立依目標讀者不同而不同的推薦清單，做個人化的新聞推薦。所提的方法結合了主題模式的技術、基於內容的過濾方法以及協同式過濾機制，藉由基於內容的過濾方法作為基礎，以及讀者對抽獎商品的喜好，來研究對於新聞推薦成效的影響。

Personalized recommendation systems have become a critical service for helping users to find items which are suitable and useful for them. However, in accordance with the change of conditions, a user’s reading interests may change over time. Hence, for online news reading, it is important to recommend articles that match each user’s dynamic preferences. Moreover, most of the existing methods obtain information only from the news read by users such as news contents, categories, and keywords. Instead of just focusing on news information, the sweepstakes-participating records were also taken to be our source data to find out users’ potential interests. A personalized news recommendation method is presented in this paper. Our source data were obtained from the website: NIUS news (www.niusnews.com), a female-oriented news website that provides news in Chinese. In order to make accurate recommendations, we adopted Latent Dirichlet Allocation (LDA), one of the topic modeling techniques to process both news contents and the descriptions of sweepstakes-items. The values of topic distributions were then used to build the user profile. Finally, by measuring the similarity between the profile of the target user and the candidate items, a personalized news recommendation is provided. The proposed method composed of topic modeling techniques, content-based filtering, and the ideals of collaborative filtering. We mainly focus on sweepstakes-item selection preference to evaluate our recommendation performance.