辨識高價值顧客:探討顧客線上行為的影響及長短期記憶模型之應用

Abstract

顧客價值衡量對於企業的顧客關係管理非常重要，透過區分不同價值的客群，企業可以個別擬定不同的行銷策略。然而隨著科技網路的快速發展，企業之間的競爭越趨激烈，因此比起拓展新客源，專注於保留並提升高價值客群的顧客價值，可以幫助企業花最少資源得到最多的利潤。此外隨著全通路( omnichannel)的概念興起，企業能蒐集到越全面的顧客資料，也可以藉由分析以了解顧客的行為模式。因此本研究欲探討顧客線上行為對於顧客價值的影響，並驗證使用行為資料後是否能提升模型預測的表現。由於行為資料非常龐大且複雜，因此本研究首先會使用2種不同的方式來對行為資料進行前處理，接著我們主要提出3種模型，並進行5種不同模型框架的實驗，最後進行模型表現比較及解釋。 本研究最終有三個結論。第一，在所有實驗模型中，僅使用單一行為事件序列進行預測的模型表現最佳。第二，我們發現顧客的下單頻率及線上互動頻率是高價值顧客預測的重要指標，兩者頻率越高代表此顧客越為重要。第三，本研究也發現若顧客有註冊、搜尋、加入購物車、結帳或是購買的行為，未來一年成為高價值顧客的機率就會提高；但相反的，若只有頻繁地查看主頁面及廣告來源導入的紀錄，該機率就會下降。

The concept of customer value is important to customer relationship management for firms. By dividing the customer base into segments based on customer value, companies could develop different marketing strategies. While with the rapid growth of the Internet, competition becomes more and more fierce. So instead of acquiring new customers, it is more effective to achieve better outcomes with less cost by only focusing on retaining and enhancing high-value members’ customer value. Also, with the rise of omnichannel business, companies could get more comprehensive customer data, and through analysis, they could understand their customers better. Consequently, this study aims to explore the influence of customer online behaviors on customer value and to confirm whether using the behavior data could improve the most valuable customers’ model performance. As the behavior data is too enormous and complex, we firstly preprocess it in two ways. Next, we primarily propose 3 types of models and conduct 5 experiments on different frameworks. At last, we compare the prediction results and interpret the outputs. In this research, there are three findings. First, only using behavior data preprocessed by the event-based method to predict the most valuable customers is the best model. Second, we observe that customers’ purchases and online interaction frequency are the most crucial features for predicting. The higher the frequency, the more important the customer is. Finally, we discover that if the customer has behaviors like searching, registering, adding to cart, checkout, or purchasing, the probability of being an important customer rises; on the contrary, if the customer only has lots of traffic source records or main page viewing, the probability decreases.