在後疫情時代，航空產業市場結構發生劇烈變遷，經營核心已由傳統的「運能導向」全面轉型為數據驅動的「精細化顧客關係管理(CRM)」。鑑於旅客數位行為模式日益複雜，傳統模型難以精準構建全方位旅客畫像。本研究旨在解決此研究困境，透過整合人口屬性(PTY)、訂位(RES)、搭機(FLN)、旅行習慣(TVL)、時空移動軌跡(OND)、商品喜好(MSL)及行銷與服務互動(PMT)等七大維度，系統性建構包含 433 個特徵值之高維特徵空間。相較於既有文獻之變數最高採用 50 個特徵變數，本研究萃取之變數維度為航空領域之最，旨在捕捉旅客生命週期早期(Early Lifecycle)微弱但關鍵的行為軌跡，為入會傾向預測奠定堅實數據基礎。本研究之核心目的在於建立具備實務解釋力與控制偽陽性投放成本之旅客入會傾向預測模型。研究採用 S 航空公司 2024 年上半年數據，經資料清洗後篩選 2，112，347 筆有效樣本作為研究母體，並結合「滑動視窗法(Sliding Window Method)」定義目標變數。針對航空數據中正負類樣本比例約 1:62 之極端類別不平衡問題，本研究採用欠採樣(Under-sampling)及 SMOTE 過採樣技術，藉此打造平衡的模型訓練數據集。實驗架構以五折交叉驗證(5-Fold Cross-Validation)，於五組子集中共產出 25 組獨立預測模型；分類演算法涵蓋邏輯斯迴歸、決策樹、梯度提升樹、類神經網路及集成學習，輸出傾向機率分數，並規劃時間外測試(Out-of-Time Test)以驗證模型之穩健性與韌性。研究結果顯示本研究成功建立一套結合多維行為特徵與機器學習技術之入會預測架構，顯著提升航空公司對潛在會員之識別能力。此架構不僅協助S 航空公司從「大眾行銷」轉型為以「精準預測分析」為核心之戰略模式，更為航空業 CRM 管理提供具高度落地價值之決策支援參考。;In the post-pandemic era， the aviation industry has witnessed a paradigm shift in market structure， pivoting its operational focus from traditional capacity-driven models to data-driven， granular Customer Relationship Management (CRM). Given the increasing digitalization and complexity of passenger behavior， conventional approaches struggle to construct a comprehensive behavioral profile. To address this challenge， this study integrates seven key dimensions—Party (PTY)， Reservation (RES)， Flown (FLN)， Travel (TVL)， Origin and Destination (OND)， Merchandise (MSL)， and Promotion and Interaction (PMT)—to systematically orchestrate a high-dimensional feature space comprising 433 unique attributes. Compared to prior research， this study extracts the highest number of variables in the aviation sector， aiming to capture subtle yet critical "behavioral fingerprints" during the early stages of the passenger lifecycle， thereby establishing a robust empirical foundation for predicting enrollment propensity.The core objective of this research is to develop a passenger membership propensity model that balances practical interpretability with the cost control of false-positive targeting.. Utilizing data from Carrier S for the first half of 2024， 2，112，347 valid records were filtered as the research population， with target variables defined through a sliding window method. To mitigate the extreme class imbalance (a positive-to-negative ratio of approximately 1:62)， this study employs under-sampling and SMOTE over-sampling techniques to construct a balanced training dataset The experimental framework utilizes 5-fold cross-validation， generating 25 independent predictive models across five sub-sets. The classifiers include Logistic Regression， Decision Trees， Gradient Boosting Machines， Neural Networks， and Ensemble Learning. A final propensity probability score is outputted via an ensemble mechanism， supplemented by an Out-of-Time (OOT) test to verify model robustness and resilience.The experimental results indicate that this study successfully establishes a membership prediction framework combining multi-dimensional behavioral features with machine learning techniques. It significantly enhances Carrier S’s capability to identify potential members， facilitating a strategic transition from mass marketing to precision predictive analytics. Furthermore， this research provides a highly actionable strategic roadmap for CRM management within the aviation industry.
