應用集群分析於精準行銷之研究-以企業軟體為例

隨著訂閱授權並交付軟體的 SaaS（Software as a Service，簡稱 SaaS）軟體即服務出現，預測模型的應用將可以為企業軟體業者提升競爭力。企業軟體業在目標客戶的預測上，常常面臨資料蒐集不易之困境。倘若能依循零售業的方式，利用資料庫中的顧客購買紀錄，作為預估未來市場的決策依據。本研究採用 RFM 指標中三項指標進行顧客價值之兩階段集群分析，再運用 CART 決策樹將客戶進行分析，建構出預測模型，進而探討各集群間的差異性。透過透過 UCI 公開資料庫的某英國批發零售商銷售總筆數 530108 之交易資料，建立預測模型，分析該企業的顧客特徵值。根據結果，給予企業軟體業者、廣告業者以及後續相關領域參考。茲將本研究重要發現分述如下：一、精準行銷與廣告策略為正相關，行銷目標在於消費者體驗上能更進階，同時降低廣告成本並創造更高的收益，最終進行付費購買。二、RFM 模型與兩階段集群分析將線上零售商客戶進行分群，從客戶變動的消費行為對其產生特徵值標籤後，將顧客分為「高消費型客戶」、「潛力型消費型客戶」、「流失型客戶」等三種類型。三、建立模型方面，使用「分類與回歸數」（Classification and Regression Tree，簡稱 CART）決策樹算法建構模型，結果發現決策樹的顯著度為 95 %，顯示決策樹能提供對應的解釋規則。

關鍵字

精準行銷； RFM 指標；集群分析； CART 決策樹

並列摘要

With the emergence of SaaS (Software as a Service) that licenses and delivers software, the application of predictive models will enhance the competitiveness of enterprise software providers. The enterprise software industry is often faced with the difficulty of data collection in the prediction of target customers. If we can follow the way of the retail industry, the customer purchase records in the database can be used as the basis for decision-making in estimating the future market. This research uses the three indicators in the RFM index to conduct a two-stage cluster analysis of customer value and then uses the CART decision tree to analyze the customer, construct a prediction model, and then explore the differences between the clusters. Based on the transaction data of a UK wholesale retailer with a total number of 530,108 sales through the UCI public database, a predictive model was established to analyze the customer characteristics of the company. According to the results, it is given to the enterprise software industry, advertising industry, and subsequent related fields for reference. The Major findings are as follows: 1. Precision marketing is positively related to advertising strategies. The marketing goal is to make the consumer experience more advanced, while reducing advertising costs and creating higher income, and finally making paid purchases. 2. The RFM model and two-stage cluster analysis group online retailer customers, and after generating eigenvalue labels from customers' changing consumption behaviors, customers are divided into "high consumption customers", "potential consumption customers", and "churning customers". 3. In terms of model building, the "Classification and Regression Tree" (CART) decision tree algorithm was used to build the model. The results showed that the significance of the decision tree was 95%, indicating that the decision tree can provide corresponding explanation rules.