網路的蓬勃發展,加上智慧型手機、平版設備等熱銷,讓資訊取得更為方便,也讓交易的方 法產生很大的變化,如何能吸引客戶的目光,讓其停留在其所提供的網頁上,提供用戶個性化的 服務,不再是所有用戶都是一樣的服務內容,因此個性化也成為電子商務中的一個熱門趨勢。然 而能否在用戶之間挖掘出有效資訊,發展出一套推薦機制,可依客戶的偏好與品味來提供其所需 要的服務內容,呈現在客戶所瀏覽頁面上,讓用戶對所推薦的內容產生興趣,進而刺激購買慾望, 提升用戶的購買意願。 本論文使用資料探勘技術建置推薦系統進行研究,主要內容包括資料變數篩選及預測模型建 立。在資料變選篩選方面,運用主成份分析(Principal Component Analysis, PCA)將變數維度縮 減,再結合奇異值分解(Singular Value Decomposition, SVD)與 k-平均法(k-means method)分 群技術對資料進行資料分析,篩選出代表性的變數。在建立模型方面,透過資料探勘演算法,包 括群集演算法(Clustering Algorithm)、關聯規則演算法(Association Algorithm),並與 PCA 結 合 SVD 形成預測模式再與傳統直接使用群集與關聯法則進行比較。 實驗資料以某網路公司的用戶折價卷瀏覽記錄為例,利用變數篩選與模型建立來架構推薦系 統之預測模型。結果發現 PCA 結合 SVD 的預測方式較直接使用關聯法則的預測方式來得有效; PCA 結合 SVD 後的群集分群更能有效找到偏好相似的用戶群,也較直接使用群集演算法的分群 效果來得明顯。
The vigorous development of the Internet, plus common usage of smartphone, tablet pc or pad equipment do make information spread more conveniently; meanwhile, it also impact ways of transactions dramatically. How to catch the eyeballs of customers, push them to stay longer on the pages is the key point. Therefore, providing users personalized service, not just one content fits all is the current trend for e-commerce. For the concept, providers wish to excavate effective information between users to develop a set of recommended mechanisms, which can allow them to provide services based on customers’ preferences and tastes, recommend certain contents to generate interests in order to stimulate the desire, even create the user's willingness to buy. This study uses data mining techniques to build recommender system for research, the main contents include data variables screening and establishing prediction model. Variable selected screening using Principal Component Analysis (PCA) to the variable dimension reduction, combined with the Singular Value Decomposition (SVD) and k-means method clustering techniques for data analysis, screening a representative variable. Modeling, through data mining algorithms, includes cluster algorithm of Association Algorithm, and combined with PCA SVD form the prediction mode cluster with association rules are compared with traditional direct use. The example used in the research is about coupon browsing history of an Internet company by using of the variables screening model to structure prediction model of the recommendation system. It is found that PCA combined SVD forecast of association rules more effective than direct use; the cluster grouping after the PCA combined SVD preferences more effective to find similar user base than directly using the cluster algorithm clustering effect is more significant.