基於評論探勘及評分因素分析之使用者喜好預測

Abstract

線上評論網站是熱門的評論資訊分享平台,提供使用者參考評論意見與評分來決定 消費行為,例如購買商品或是到訪商家。然而大量的線上評論資訊造成資訊過載問題, 使用者不容易找尋符合其喜好的商品或商家,因此分析線上評論網站之評論與評分來預 測使用者之喜好評分,並進行個人化推薦為重要之研究議題。 傳統使用者喜好評分預測方法大多以協同過濾方法分析歷史評分紀錄來預測個別 喜好評分,然而使用者之喜好評分通常受到不同面向的評分因素所影響,使用者對於各 面向有不同的喜好重視程度,而不同商家雖然有不同面向的表現,仍然可能獲得不同面 向喜好使用者所給予之相似評分。因此,僅分析使用者給予商家的評分來進行預測,無 法有效分析使用者在各面向之相似喜好以及商家在各面向之相似表現,將導致喜好預測 上的誤差。因此,傳統僅以使用者評分進行喜好分析預測之方法,有其限制而無法有效 預測喜好評分。 本研究提出新的使用者喜好評分預測方法,考量基於不同面向的評分因素包括使用 者對於不同面向的喜好重視以及商家在不同面向的表現,本研究探勘評論文字來發掘各 面向的意見語意,並以此為基礎分析使用者評分,發掘不同面向的評分因素,建立基於 面向喜好重視之使用者喜好評分預測模型,及商家的面向表現模型,進而預測使用者對 不同商家之喜好評分。本研究收集線上評論網站 Yelp 之資料進行實驗評估,實驗結果 顯示,本研究所提的方法優於傳統使用者喜好評分預測方法,能改善評分預測的準確性。

Online review websites are nowadays popular information sharing platforms, which help users decide whether to buy products or visit business stores by referring the review opinions and ratings. However, a large amount of review information results in information overload problems and difficulty for users to find preferred products or business stores. Accordingly, it is an important issue to predict user preferences and make recommendations by analyzing the review opinions and ratings on the websites. Traditional rating prediction methods usually adopt collaborative filtering to predict user ratings based on historical rating records. However, users’ preference ratings are usually affected by the aspect-based ratings factors including user preference emphases and business performances on various aspects. Specifically, different users may have different emphases on aspect preferences. Business stores with different aspect performances may receive similar ratings from users with different aspect preferences. Consequently, predicting user preferences by only considering user ratings of business stores, cannot effectively identifying users with similar aspect preferences and business stores with similar business performances, and thus may result in poor predictions. Traditional methods, which only consider historical user ratings, are limited and not effective in predicting user ratings. This research proposes a novel rating prediction method considering the aspect-based ratings factors. First, the review texts are analyzed to extract the opinion semantics of various aspects. Second, user ratings on aspect semantics are analyzed to discover the aspect-based rating factors, which are used to build the user rating prediction model and business performance model. Third, the two models are then used to predict user preference ratings on business stores. Finally, experiments are conducted to evaluate the proposed method using Yelp dataset. The experiment results show that the proposed method outperforms traditional methods and can improve the accuracy of rating predictions.