萃取線上評論品質特徵的遊戲類行動軟體推薦系統開發

隨著行動科技的進步與成熟，民眾擁有智慧型手機的現象變得普遍。智慧型手機除了提供通訊服務外，也衍生許多讓使用者的生活更為方便的行動軟體。在下載特定軟體前，使用者通常會先瀏覽該行動軟體的公開資訊。為了協助使用者做出下載決定，APPLE及Google均有其推薦機制。其推薦機制主要是參考使用者下載的行動軟體紀錄，將下載同樣行動軟體的其他使用者還下載過哪些其他的行動軟體作為推薦標的。這樣的推薦方式雖然有其效用，但是未考量使用者對特定的品質偏好，使得推薦結果未必能滿足使用者的喜好。本研究基於上述提到的缺點進行改進，開發出遊戲類行動軟體推薦系統。本研究的作法如下：盡可能收集使用者撰寫的線上評論，參考遊戲相關的體驗品質與品質特徵後歸納評論中的品質詞，統計每一位使用者在其評論中提到的各項品質詞出現次數並轉換成使用者的品質特徵偏好向量，接著以此向量為所有使用者進行K-means分群。根據分群，本研究對特定使用者以其所在的群中被其他使用者高度評價的遊戲作為推薦標的。本研究以實驗方式使用成對樣本t檢定將本研究的推薦方法與現行推薦方法進行效能比較，結果顯示在準確率與召回率方面，本研究的推薦方法顯著優於現行的推薦方法。

關鍵字

推薦系統；線上評論；體驗品質； K-means

並列摘要

With the advancement and maturity of mobile technology, the phenomenon of people owning smartphones has become common. In addition to providing communication services, smartphones also spawn many mobile software that make users' lives more convenient. Before downloading the software, users usually browse the public information of the mobile software. In order to assist users in making download decisions, both APPLE and Google have their own recommendation mechanisms. The recommendation mechanism is mainly to refer to the user's downloaded mobile software records, and to recommend other mobile software that other users who have downloaded the same mobile software have downloaded as the target of recommendation. Although such a recommendation method has its utility, it does not consider the user's specific quality preference, so that the recommendation result may not be able to satisfy the user's preference. This study is based on the above-mentioned shortcomings to improve, to develop a recommendation system for mobile games. The method of this study is as follows: Collect online reviews written by users as much as possible, refer to game-related Quality of Experience and Quality Characteristics to summarize the quality words in the reviews, and count the number of quality words mentioned by each user in their reviews and convert them into the user’s quality feature preference vector, and use this vector to perform K-means grouping for all users. According to the grouping, this study recommends games that are highly valued by other users in the group of specific users as the target of recommendation. This study uses paired sample t-test in an experiment survey to compare the effectiveness of the recommended methods in this study with the current recommended methods. The results show that the precision and recall, the recommendation method of this study is significantly better than the current recommendation method.