線上媒體平台推薦系統之研究

Abstract

線上媒體服務不斷的興起，提供各種面向的服務給所有的使用者，包括新聞、音樂、影音、電子商務、遊戲、以及虛擬世界等等。這些線上媒體平台吸引各類型的用戶進入，產生流量。並以此為基礎，透過會員付費或是廣告販售將流量變現，藉此產生營收。因此流量對於線上媒體平台而言，是關乎生存與獲利的重要因素。而線上流量的產生很重要的一部分來自內部橫向串連，透過有效的介面設計、動線規劃、以及相關內容的推薦去增加讀者在站內的互動、閱讀、與商品瀏覽等等來提升網站流量，創造收益。 本研究針對線上媒體平台之內容推薦進行探討，包含以社交互動為基礎的虛擬世界與以內容為基礎的新聞平台。虛擬世界使用電腦環境模擬現實世界，會員在虛擬世界中透過虛擬人物與其他會員進行互動。其最重要的營收來源為虛擬商品的販售，有效的商品推薦將能正面拉抬整體營收。本研究對虛擬世界的社交行為影響力進行分析，了解影響力的產生與互動行為的關係，並提出一個基於社交行為的虛擬商品推薦方法，並以線上虛擬世界：嚕米玩樂城市之實驗資料進行實驗評估。實驗結果顯示，考慮社交行為的虛擬商品推薦方法優於傳統的商品推薦方法。 而線上新聞平台通常提供多種面向的內容來吸引不同類型的讀者，讓使用者在站內流動，產生額外的流量。妞新聞網站長期與品牌廠商合作舉辦贈獎活動，提供贈品吸引使用者參加，傳遞品牌價值，而新聞平台則獲得使用者黏性與流量。若能夠有效的對使用者推薦活動，將能夠對網站流量產生正向的提升。本研究分析使用者的新聞閱讀喜好特徵及活動參與，並建立新聞與活動的關聯性，提出使用者瀏覽文章時之線上活動推薦方法，並整合推薦清單替換策略，動態調整推薦清單以獲得更好的推薦結果。 本研究依據提出之推薦方法，在妞新聞平台上實作一套可因應需求動態調整的線上推薦系統。此系統採用REST架構來執行與妞新聞網站之資料溝通，於每日凌晨進行資料匯入與資料處理，並透過動態服務器資源調配於有限時間內進行使用者喜好分析與離線推薦結果計算，在運算需求與成本上達到平衡。並於讀者進入新聞頁面時，即時整合預先運算完成的模式庫進行線上分析與推薦，最後依據此線上系統進行線上推薦實驗驗證，實驗結果顯示，本研究所提方法能夠提供更有效的推薦，研究成果將能為網站帶來更高的流量，進而提高網站獲利。

Online media platforms have been increased tremendously in recent years. These online services gain website traffic from all around the world, and monetize their traffic by ads, member fees, or e-commerce. A good method to increase their traffic is definitely crucial to uplift their revenue. Website traffic mainly comes from in-site page views, which are affected by user interface, user experience and the information providing. Providing right information in the right place could attract user’s eyeball, and encourage them to view more contents, or buy more goods. This research mainly investigates the recommendations for online media platforms, including the social-based platforms (e.g. virtual worlds) and the content-based media platforms (e.g. news platforms). Virtual worlds (VW) are computer-simulated environments which allow users to interact with other people through their own virtual characters. In VWs, virtual goods are the most important revenue resource. Different from traditional product recommendations, users’ buying behaviors in VWs are highly affected by their social activities. This work proposes a social-based recommendation approach to recommend virtual goods that considers the users’ virtual life features and the contact influences derived from the interactions with social neighbors. The research highlights the importance of social interactions in virtual goods recommendation. The experimental evaluations are conducted by using the data retrieved from an on-line VW platform: RooMi. The experiment results show that the proposed method, considering social interactions and social life circle, can achieve better performance than existing recommendation methods. The content-based news platforms always provide multi-lateral content to attract users through their most interested topics. Combining online news websites with attending activities can attract more users and create more benefit. This study designs novel recommendation methods to predict the long-term and online user preferences on activities based on the association analysis on users’ browsing news and attending activities. Furthermore, novel recommendation adjustment strategies are proposed to dynamically adjust the recommendation list to increase users’ click-through rates. Finally, an online recommender system is implemented on an online media platform: Niusnews. The online recommender system is dynamically adaptive to cost and efficiency. Online experimental evaluations are conducted on the news platform. The result of the online recommender system demonstrates that our proposed approach can enhance the effectiveness of recommendations.