SARS－結合朋友主動資訊之推薦機制

Abstract

推薦系統的研究已行之有年，主要以內容導向過濾法、協同式過濾法與混合式過濾法等為基本分類。近來由於社群網路的興盛，許多研究紛紛開始注意到，加入社會影響的考量，如最具代表性的信任機制，在面對冷起始的情況下，更能顯著提升推薦效能。 對協同式推薦系統來說，建造使用者輪廓的資訊，其來源從使用者本身提供的喜好評分資訊，透過這些喜好與他人的喜好相似度，預測使用者可能會喜歡的商品；信任機制方面，則是利用非直接的管道，設想使用者與高度信任的對象在喜好興趣方面相似的可能性較高，歸納而言，主要的研究精神皆在於「如何找尋更多與使用者喜好正相關的推薦依據」，故本研究希望能從「如何拓展直接掌握使用者興趣喜好之來源」的另一種角度切入，找出可以再提升推薦效能的方法。 為此本研究期望透過近來社群網路興盛之便，利用社群網路的交友機制，蒐集使用者朋友對使用者的主動推薦商品資訊，讓系統除了能從使用者本身商品評分外，亦可由朋友的主動推薦掌握使用者可能會有的興趣，是故本研究提出Social-based Adjustment Recommender System(SARS)演算法，將朋友對某商品的主動推薦次數，融入傳統協同式User-based演算法計算出的預測分數中，將其作為最終推薦預測結果。實驗結果證實，利用使用者朋友的主動推薦作為推薦系統預測依據之一，確實能達到推薦預測準確率提升效果。

Recommender system has been studied for many years. The basic types of recommender system are: content-based, collaborative filtering, and hybrid approach. Recently, with the prosperity of social network, many studies begin to notice that the recommendation accuracy can be enhanced by taking social influence such as trust mechanism into consideration, especially when there is a cold start problem. For collaborative filtering recommender system, the source to construct user profile comes from rating information based on user's personal preference. The items we recommend to a person are based on the similarity between preferences of users. As for trust mechanism, it is assumed that the similarity between users' preferences positively correlate with their trust relationship. In sum, the main idea is focus on finding a recommender support which has positive relative with users’ preferences. Therefore, our work aims to enhanced recommendation accuracy by increasing the source of users’ preferences that system can directly get. Our study tries to use social network’s friends-making mechanism to collect user’s interest through user’s friends’ active opinions. Recommender system can analyze user’s preference not only from user’s ratings, but also from friends’ active recommendations. So we propose social-based adjustment recommender system (SARS) algorithm, which merged the counts of friends’ active recommendations into the score predicted by user-based collaborative filtering as final predict score. According to experiment result, we prove that recommendation accuracy can be improved by considering friends’ active opinions.