運用點擊流之 Signal RFM 與 Signal CAI 分析顧客價值

Abstract

本研究之目的為通過點擊流數據分析電商顧客價值。我們首先應用海盜指標 (AARRR) 並將目標放在已活躍的顧客上，而後定義 Future Return 和 Future RFM 這兩項任務以協助我們進行價值衡量。在特徵工程中，我們將特徵分為五 大群，分別為 RFM 特徵群，CAI 特徵群，矩陣特徵群，偏好特徵群和其他特徵 群。其中特別的是，我們將行為視為某種價值信號，並藉此定義 Signal RFM 和 Signal CAI 作為技術指標來幫助電商快速有效地衡量顧客價值。我們訓練了邏輯 回歸 / 線性回歸、XGBoost、前饋神經網絡和 LSTM-CNN 模型來預測 Future Return 和 Future RFM。在實驗中，XGBoost 展現了其強大的效能和優秀的預測 能力，並被視為最佳單一模型。而 LSTM-CNN 通過採用最近顧客行為資料作為 模型附加輸入，顯示了其擊敗 XGBoost 的潛能。最後，我們再透過簡單的線性 回歸來集成學習所有單一模型的預測，並以此達到最佳結果。

The purpose of this research is to analyze E-commerce customer value through clickstream data. At first, we apply AARRR model and focus on active customers. Then we define Future Return and Future RFM tasks to help us estimate the value of customer. For feature engineering, we generate five types of features: RFM, CAI, Matrix, Preference and Other. Especially, we consider behavior meaning as some kind of value signal and formulate Signal RFM and Signal CAI. These two technical indicators can help E-commerce companies measure customer value efficiently. Furthermore, we train Logistic / Linear Regression, XGBoost, Feed Forward Neural Network and LSTM-CNN models to predict Future Return and Future RFM. XGBoost has presented its outstanding performance and is considered to be the best single model for our tasks while LSTM-CNN shows the potential to beat XGBoost by adopting recent customer logs as additive model inputs. The ensemble of all the models by a simple Linear Regression model help us reach the best performance in the end.