| 摘要: | 在疫情蔓延全球期間，由於各地區實施封鎖措施部分實體店面被迫關閉，使得消費者的購物方式迅速從線下轉向線上，這種轉變加速了電子商務的發展，也促使消費者支付方式的改變，其中先買後付(Buy Now Pay Later)服務快速崛起。先買後付為消費者提供了可遞延或分期付款的支付方式，消費者無需即時支付全額，降低消費者經濟壓力，同時增加購買意願。特別是在經濟不確定的後疫情時代，此外，這種付款方式在心理上也為消費者減輕支付時的壓力。
由於先買後付服務在全球迅速普及，對電子商務平台而言，能夠準確識別和預測潛在使用此服務的消費者並推播其使用，期能藉此降低消費者支付透明度以提升購物衝動，有助於提升消費者購買意願、消費金額及消費者整體滿意度，故本研究以機器學習技術建立先買後付消費者預測模型，以提供電商平台更有效地制定行銷策略和精準推廣增加服務使用率。
以電子商務平台消費者交易資料，包括購買頻率、金額、商品偏好及支付偏好等，鑑於原始資料集中存在的資料不平衡問題，本研究採用了六種資料不平衡技術，如隨機過採樣、隨機欠採樣、SMOTE、ADASYN、ENN及SMOTE+ENN技術，以提高模型對少數樣本的預測能力。在模型建立階段，選擇搭配六種機器學習，包括邏輯迴歸、隨機森林、梯度提升技術、KNN、樸素貝葉斯、神經網路進行模型建立，並採用十折交叉驗證方法來評估各個模型的結果。研究結果顯示，比較不同資料不平衡及機器學習模型，資料不平衡經SMOTE+ENN處理並應用梯度提升技術，模型在整體預測表現及穩定性結果最佳。
;During the pandemic, lockdown measures forced many stores to close, rapidly shifting consumer shopping from offline to online. This shift not only accelerated the development of e-commerce but also led to changes in consumer payment methods. Buy Now, Pay Later service emerged rapidly, offering consumers a way to defer or installment payments. This payment method does not require consumers to pay the full amount immediately, reducing their financial stress and increasing their willingness to purchase, especially in the economically uncertain post-pandemic era. Additionally, this payment method psychologically reduces the stress of payment at the time of purchase, thereby enhancing the shopping experience and consumer satisfaction.
Due to the rapid global growth in the usage of Buy Now, Pay Later service, it is crucial for e-commerce website to be able to accurately identify and predict potential users of this service and promote its use. By reducing payment transparency to increase impulse purchase, which can help increase customer transaction value and consumer satisfaction. Therefore, this study uses machine learning techniques to establish a predictive model for Buy Now, Pay Later consumers to provide e-commerce website with more effective promotion and precision marketing to increase service usage rates.
The study utilizes consumer transaction data from e-commerce website, including purchase frequency, amount, product preferences, and payment preferences. Given the class imbalance problem in the original dataset, the research employs six data imbalance techniques, such as Random Over Sampling, Random Under Sampling, SMOTE, ADASYN, ENN, and SMOTE+ENN, to improve the model prediction ability for minority samples. During the model building phase, six types of machine learning, including Logistic Regression, Random Forest, Gradient Boosting, KNN, Naive Bayes, and Neural Networks, along with 10-fold cross-validation methods to evaluate the results of each model. Comparing different data imbalance techniques and machine learning models, the model preprocessed with SMOTE+ENN and applied with Gradient Boosting performs best overall and in terms of stability. |
