基於協同主題模型與主題熱門分析之個人化新聞推薦

Abstract

網路上閱讀新聞愈趨普及，然而由於資訊爆炸，使用者面對海量新聞文章，很難精確地找到有興趣的文章。娛樂新聞網站大多推薦最新新聞、或熱門新聞，較少針對使用者喜好進行推薦。相關研究提出協同主題模型方法，透過隱含狄利克雷分布方法找出使用者的隱含主題喜好及文章的隱含主題特徵，以此為基礎進而以矩陣分解法推導使用者及文章之隱含主題特徵並進行推薦。文章主題之熱門程度及名人影響力會影響使用者之喜好。然而以協同主題模型為基礎的相關研究並未考量主題熱門分析及名人影響力。本研究提出基於協同主題模型與主題熱門分析之新推薦方法，所提方法整合LDA隱含主題分析、主題熱門分數及名人影響力進行矩陣分解及推薦。本研究以娛樂新聞網站資料進行實驗評估，實驗結果顯示，所提方法能有效提升新聞推薦之精確性。

Reading news on the internet is getting prevalent. However, due to the information overload problem, it’s difficult for users to find their interested articles. The online entertainment news website usually recommend latest or popular news to users. Few would recommend news articles to users based on their preferences.

Existing studies have proposed the Collaborative Topic Modeling, CTM method, which combines Latent Dirichlet Allocation and Matrix Factorization to recommend articles. The CTM method uses LDA to extract latent topic features of articles, and then adopts MF to derive the latent factors of users and articles based on the LDA latent topic features. The CTM methods have been shown to provide more accurate recommendation. However, existing CTM methods do not consider the popularity of topics and celebrity influence, which may affect users’ preference on articles.

This study proposes a novel recommendation approach, which enhances CTM methods by considering topic popularity scores and celebrity influence to improve the recommendation quality. This study uses a data set collected from a popular entertainment news website to evaluate the proposed approach. The experiment result shows that the proposed approach can achieve more accurate recommendation than existing CTM methods.