實作HITS演算法於實況串流頻道推薦系統

Abstract

實況串流頻道服務近年來蓬勃發展, 各式各樣的頻道如雨後春筍般產生，播報主藉由撥 放他們的視頻來吸引廣大的收視群眾，而觀眾也藉由這個平台來跟播報主互動，並和 其他觀眾一起觀看同樂。然而在這個平台上，每天都會有數以百計的視頻上傳，平台 本身的推薦系統並不是那麼的有效率，例如他們推薦的內容是以最多人觀看的影片來 做推薦，而沒有根據個人的觀看行為來做調適。為了要改良現有低推薦率的情況，我 們提出了hybrid HITS algorithm for Channel Recommendation (HyHITS_CR)。首先先改良 HITS algorithm，這是為了實況串流頻道服務上追隨者與被追隨者關係而調整。此外， 在我們的觀察中也發現，很多觀看者只會鎖定在某幾個特別的頻道，因此我們把這個 因素考量進我們的HITS_CR 演算法中，進而成為HyHITS_CR 。在此篇論文中，我們 把提出的方法套用在知名的實況串流頻道平台-Twitch 上面，而實驗結果也如我們預期 比其他的演算法的效能好，HyHITS_CR 推薦成功率比現有的方法高了10 個百分比。

Live streaming service has been developed prosperously in recent years, and the platforms provided by it are full of various types of channels. Broadcasters broadcast their videos to attract large numbers of viewers to watch. The service also provides a platform for viewers gathering together to watch channels and interact with others. However, over hundreds of videos uploaded everyday, the recommender system provided by existing platform seems to be ineffective, such as applying most viewed algorithm for recommendation. In order to improve the low accuracy of existing recommender system, we propose the hybrid HITS algorithm channel recommendation (hyHITS_CR). First, we modified HITS algorithm for follower-followee relationship on live streaming channel. Besides, we also find some viewers may focus on few particular channels. Thus, we consider this feature into our method to become hybrid HITS algorithm. In this paper, we apply the proposed method on Twitch - the most well-known live streaming service, and the result outperforms others in our experiment. The hit rate of hyHITS_CR is 10% higher