隨著生活方式和環境的改變以及新冠肺炎(COVID-19)的影響，驅使消費者更加注重自身的日常健康，對於保健食品的需求變得更為迫切。在保健食品產業中，電話行銷是主要的銷售通路之一，如何從大量的名單中找到最適合的顧客進行推銷成為電話行銷人員關注的課題。本研究基於現有文獻對信任的定義建立操作型定義，以資料驅動的方式透過提取公司系統中的資料來衡量信任指標，作為預測再購模型的輸入特徵，並建立了兩種具有注意力機制的LSTM模型來預測顧客是否會再次購買。 由於資料不平衡是預測顧客是否再購的常見問題，這種不平衡的情況會對模型預測效能產生影響，因此本研究分別針對原始不平衡的資料以及平衡資料來進行顧客是否會再購之預測，以更全面地評估模型之效能。第一種模型是 LSTM-SA，該模型為LSTM融合自注意力機制，透過自注意力機制來找出每位顧客購買模式中的時序特色，該模型實驗結果在原始資料下F1-score可達48.9%、AUC可達92.4%，在平衡資料下F1-score可達89.9%、AUC可達91.8%。第二種模型是KVMN-LSTM，該模型為Key-value 記憶網路結合LSTM，透過Key-value 記憶網路中的注意力機制來找到與欲預測顧客最相似的其他顧客，再根據這些相似顧客的購買行為來預測該顧客的購買行為，該模型實驗結果在原始資料下F1-score可達49.2%、AUC可達93.9%，在平衡資料下F1-score可達89.2%、AUC可達92.4%。;With the changes in lifestyle, environment, and the impact of COVID-19, consumers have become more focused on their daily health, leading to an increased demand for health supplements. In the health supplement industry, telemarketing is one of the main marketing channels, and identifying the most suitable customers for sales from a large list has become a crucial concern for telemarketers. This study establishes operational definitions of trust based on existing literature and uses a data-driven approach to measure trust indicators by extracting data from the company′s systems. These indicators serve as input features for predicting customer repurchase behavior, and two LSTM models with attention mechanisms were developed to forecast whether customers will make repeat purchases.
Data imbalance is a common issue in predicting customer repurchase behavior, and this imbalance can affect the predictive performance of the models. Therefore, this study separately analyzes the prediction of customer repurchase behavior using both the original imbalanced data and balanced data to comprehensively evaluate the model′s effectiveness. The first model is LSTM-SA, which combines LSTM with a self-attention mechanism to identify temporal patterns in each customer′s purchasing behavior. The experimental results show an F1-score of 48.9% and an AUC of 92.4% for the original data, and an F1-score of 89.9% and an AUC of 91.8% for the balanced data. The second model is KVMN-LSTM, which combines Key-value Memory Networks with LSTM and utilizes the attention mechanism in Key-value Memory Networks to find other customers who are most similar to the target customer. Based on the purchasing behavior of these similar customers, the model predicts the target customer′s purchasing behavior. The experimental results show an F1-score of 49.2% and an AUC of 93.9% for the original data, and an F1-score of 89.2% and an AUC of 92.4% for the balanced data.
