一般認為使用者的點擊流 (clickstream) 可以代表使用者的線上瀏覽行為，然而，我們發現點擊流只能概略表示使用者的部份行為，例如：分頁切換、視窗切換等介面間的瀏覽行為因為沒有產生與伺服器的互動，所以不會出現在點擊流或日誌 (log) 中，但使用者仍然在瀏覽網頁。本文將這些行為收集並命名為「擴展點擊流」(extended clickstream)。透過建設完整的系統服務並招募受試者來同步蒐集點擊流和擴展點擊流，並對兩者進行比較分析及建構深度學習模型。我們使用含有 GRU 元件的深度學習模型，對點擊流和擴展點擊流這類型的時序資料進行「使用者下次會去什麼類型的網站」、「下次點擊會間隔多久」的多目標預測。實驗結果顯示：融合點擊流和擴展點擊流可以增進預測效能。除此之外，本文發現點擊流會因為部分網站的運作機制而多計入了使用者沒有意圖執行的行為；另外，我們也可以透過融合點擊流及擴展點擊流來區分出來自不同裝置的單一使用者;Nowadays, people often use clickstream to represent the behavior of online users. However, we found that clickstream only represents part of users′ browsing behaviors. For instance, clickstream does not include tab switching and browser window switching. We collect these kinds of behaviors and named as ``extended clickstream". This thesis builds a service to capture both of clickstream and extended clickstream, also provides an analysis of the differences between above. We use a Multi-Task learning model with GRU components to perform multi-objective predictions of ``what kind of website the user will go next time" and ``how long the interval of clicks will be" for the time series of clickstreams and extended clickstreams. Our experimental results show that combining clickstream and extended clickstream can improve the prediction performance. In addition, this article finds that the clickstream will record unintended clicks due to the operation mechanism of certain websites. Moreover, we can differentiate the single user from several devices by combining the clickstream and extended clickstream.
