重訪預測：透過多任務學習整合在線評論和管理回應

Abstract

在競爭日益激烈的環境中，顧客保留對於維持市場份額和企業持續成功變得至關重要。除了顧客資料和交易記錄外，分析用戶生成內容（UGC）能更深入了解顧客的體驗。有效的網絡護理策略也顯著影響顧客保留。有效的護理策略包含企業透過積極回覆顧客問題，並與客戶進行更多交流來建立信任感。顧客對於不同的回應方式會有不同的感受，因此企業在回覆時，除了要回覆，更要考慮如何回覆才能讓客戶滿意。 本文著重於旅遊業中的客户保留，特别是重訪行為。在預測顧客重訪時，我們使用了來自顧客評論和企業回應的文本數據。傳統方法依賴基於詞典或詞嵌入來獲得文本表徵。隨著自然語言處理技術的進步，我們現在採用上下文嵌入來更深入地理解文本語境。以往的顧客保留預測研究通常集中於單一任務，忽視了相關資訊的影響。我們採用了多任務學習方法，結合顧客評論評分作為整體滿意度和企業的回應策略，以提高預測準確性。 通過實驗，我們提出的模型PROMO在預測顧客保留以增強客戶關係管理（CRM）策略方面表現出色。結合輔助任務能夠進一步提高了預測準確性。我們的研究提出了一個有效利用來自顧客和企業的文本資料的深度學習模型。利用上下文嵌入獲得文本表徵並結合輔助任務，我們實現了更精確的預測。這項研究的實際應用意義重大，為CRM策略提供了寶貴的見解，幫助管理層做適當的回應並培養顧客忠誠度。

In a competitive landscape, customer retention is crucial for maintaining market share and sustained success. Analyzing user-generated content (UGC) provides deeper insights into consumer experiences beyond traditional consumer profiles and transaction records. Effective webcare strategies, including responding to customer inquiries and engaging in interactions, significantly impact retention, as customers react differently to various response strategies. This paper focuses on customer retention in the tourism industry, namely revisit behavior. We used textual data from customer reviews and business responses for prediction. Traditional text representation methods relied on lexicon-based or word embeddings, but advancements in natural language processing now allow for contextual embeddings that offer a deeper understanding. Past research often focused on a single task, neglecting related information. Our multi-task learning approach incorporates review ratings and response strategies to enhance prediction accuracy. Through experiments, our proposed model, PROMO, demonstrated superior performance in predicting revisits. The practical implications of our research are significant, providing valuable insights for CRM strategies and aiding managerial decisions to deliver appropriate responses and foster customer loyalty.