近年隱私權意識高漲,歐盟祭出GDPR法規,強化個資當事人權益,Google宣布將取消支援第三方cookies,行銷科技受到限縮。然而電子商務市場持續成長,電商網站必須在隱私權的限制下,追蹤使用者行為,尋求任何可以改善使用者體驗的可能。因此,本研究捨棄所有使用者特徵並去識別化,僅蒐集用戶與網站內各區塊間的互動,開發使用者行為分析系統。相較其他系統以大量使用者個資做為分析依據,本研究將使用者的活動流進行瀏覽歷程識別,並將之視為圖形資料以進行圖分析(graph analytics)。我們使用圖形相似性為基礎的多種連結預測(link prediction)方法與圖神經網路(graph neural networks)分析使用者與網站元件的互動。對外提供消費者即時商品推薦,對內則提供網站架構設計與頁面布局重構建議。本研究蒐集實際電商網站之使用者紀錄進行實驗,結果顯示透過連結預測演算法進行商品推薦,Top5推薦可得43.40%的平均命中率。利用圖神經網路進行區塊分析,預測區塊間的關聯性,Top5推薦可得72.05%的平均命中率。我們接著提出濾除「雜訊操作」的方法,可將Top5推薦的平均命中率再提升4.06%。
In recent years, the awareness of privacy is rising where the EU proposed GDPR to strengthen the rights of individual data parties, and Google announced to cancel the support for third-party cookies. Conventional MarTech(marketing technology) is restricted. However, the e-commerce continues to grow, and e-commerce websites must track user behavior and seek any possibility to improve the user experience within the constraints of privacy. Therefore, this thesis discards all user characteristics and de-identifies them, and only collects the interaction between users and various blocks in the website. Compared with other systems that use a large amount of users’ personal data as the basis for analysis, this thesis did session identification on users’ activity streams they were treated as graph data and uses a variety of link prediction algorithms based on graph similarity and graph neural networks to analyze user interactions with website elements. It provides consumers with real-time product recommendations, and provides website architecture design and page layout reconstruction suggestions to website maintainers. This thesis collects user records of actual online e-commerce websites and conducts experiments. The results show that through one of the link prediction methods for product recommendation, the Top5 recommendation can achieve an average hit rate of 43.40%. Using graph neural network to analyze blocks and predict the correlation between blocks, the average hit rate of the Top5 recommendation is 72.05%. Then, We proposed a method to filter out "noise operations", which can increase the average hit rate of the Top5 recommendation by 4.06%.