推薦系統之觀眾行為模式分析: 以臺大開放式課程網為例

Abstract

早在2010年時，推薦系統就已不再是新穎的話題，而是所有電商、串流平台甚至是餐飲業者邁向成功的一項基石。「你可能喜歡…」、「你可能想看…」、「和你瀏覽相同商品的人也…」，在現在的生活中無處沒有推薦系統，從收聽音樂到觀看電影，從網購商品至出遊旅行，生活中大大小小的決定都有著推薦列表輔助我們進行決策，就好像不管想做什麼事情、想去哪裡都有人引領我們一般，而這種便利的生活模式也早已被視作理所當然。 本次研究對象「臺大開放式課程(NTU OpenCourseWare)」缺乏一套完善且有效之推薦系統，為此本研究從其網站之使用者與課程資料做資料分析，從近60萬筆瀏覽紀錄中找出潛在的行為模式，同時考量課程本身可能出現的週期性、被觀看次序性，讓今後上線之推薦系統獲得更加全面性的優化，納入時間與序列性等參考指標，輔以推薦模型之邏輯設計與建置。 研究結果顯示，於課程資料中，除了每個課程都有自己的觀看週期性外，部分課程與課程之間亦擁有強大關聯性，時常是觀眾在此平台上一併觀看之組合，同時也存在明顯之學習次序性，觀眾在先看後看的行為上有一致的表現。此外，於用戶資料中，能夠發掘用戶存在「穩定」、「密集」點擊兩種行為模式，且密集點擊行為人也對於點擊不同課程之意願較穩定點擊行為人還高，在學習次序上也擁有著比穩定點擊行為人更加一致的學習狀態。

As early as 2010, recommendation systems are no longer new but a cornerstone of success for all e-commerce, streaming platforms, and restaurants. "You may like...", "You may want to watch..." and "People who browse the same products as you also..." are recommended everywhere in our daily life, from online shopping to movie watching, from music listening to traveling. There have great recommendation systems for introductions and suggestions for every single decision of life. No matter what we want to do or where we want to go, recommendation systems always help us make decisions. This convenience has been around us for a long time, and we cannot live without it. NTU OpenCourseWare lacks a complete and adequate recommendation system. For this reason, this research analyzes NTU OpenCourseWare's user and course data. Find potential behavior patterns from nearly 600,000 browsing records, and also consider the possible periodicity and viewing order of the course itself, so that the recommendation system launched in the future can be more comprehensively optimized. The research results show that, in course data, in addition to each course having its viewing cycle, some courses also strongly correlate with each other. It’s often a combination of audiences watching together on this platform, and they also have some apparent learning order, the audience has a consistent performance in the behavior of watching first and then watching; in the user data, it can be found that users have two behavior patterns of "stable" and "intensive" clicks, and the intensive clickers also have more willingness of watching different courses than that of the stable clickers, and their learning orders also are more consistent than the stable clickers.