以遊戲影片為例探討藉由使用者認知之相似因素改善手機遊戲推薦系統

Abstract

由於智慧型手機和平板的普及，行動平台上的遊戲成為成長最快速的產業之一。 僅僅在蘋果的 App Store 上面的遊戲就已經超過二十萬個了， 可以想見加上其他行動平台的遊戲，現在的行動遊戲數量有多驚人。 但是如同中央研究院資訊科學研究所的楊先生與他的夥伴們的研究論文所說的， 使用者很難用現有的系統找到符合他們需求的遊戲，尤其是和某個特定遊戲相似的遊戲。 本篇論文從現有的遊戲系統蒐集了 3456 個遊戲配對， 給 286 個使用者標記相似度之後，發現 Google Play 推薦的相似遊戲中， 平均有超過 80% 的遊戲被使用者認為是不相似的。 因此我們以質化訪談整理出使用者判定遊戲相似的依據， 並且提出根據這些相似因素可以大幅提昇相似遊戲推薦的準確率。 在本篇論文中，我們以從遊戲影片中擷取遊戲中的 motion vector 為例實作了一個相似遊戲推薦系統， 並且將準確度提昇了高達 42%。

Mobile gaming has become one of the fast growing industries due to the rise of the mobile market. The number of mobile games has increased to over 200 thousands on App Store, not to mention there are still plenty of games on other platforms. However, as mentioned by Yang et al. in their work, current systems fail to provide sufficient support for users to find specific games that meet their needs, especially for similar games. The problem is also supported by our qualitative study and evaluation result. According to our similarity rankings collected from 286 people, current recommender system for similar games produces over 80% dissimilar games on average. We present a novel idea that by considering users definition of similarity between games, we can significantly improve recommendations for similar games by 42%. We demonstrate our idea by implementing a recommender system that incorporates motion features of gameplay video, which is one important similarity factor mentioned by our participants in the qualitative study. Our results show that by considering only gameplay motion features, our system can generate recommendations that outperform existing recommendations for similar games.