個人化推薦系統之模型解釋與聚類分析: 以葡萄酒推薦為例

Abstract

在數位化迅速發展的現今，購買葡萄酒已不再僅限於實體通路，消費者也可以透過網路平台購買各類酒款。然而，葡萄酒作為具高度個人偏好和需要專業知識的商品，加上網路資訊的爆炸性增長，使消費者在網路上尋找合適的葡萄酒產品變得困難。因此，為了提升消費者體驗和增加銷售，建立個人化推薦系統變得尤為重要。

本研究使用一葡萄酒電商平台的銷售數據，採用推薦系統中著名的矩陣分解模型 (Matrix Factorization)，並結合貝氏個人化排序演算法 (Bayesian Personalized Ranking) 為電商平台建立葡萄酒推薦系統。然而，該模型雖能達到良好的預測效能，但模型產生的推薦結果卻難以被解釋，使得公司或用戶可能無法信任模型的推薦。

為解決這項問題，本研究引入偏最小平方迴歸 (Partial Least Squares Regression)，從葡萄酒標籤屬性的角度來分析模型的推薦邏輯。此外，使用階層式分群法 (Hierarchical Clustering) 對模型產生之酒款潛在因子進行分群，並針對分群結果探討哪些葡萄酒屬性為形成集群的關鍵。透過以上分析來為推薦結果提供合理解 釋，不僅能增加電商平台對推薦的信任度，也能深入了解平台用戶的消費行為，為推薦策略提供重要參考及依據。

In the rapidly evolving digital era, buying wine is no longer limited to physical stores. Consumers can also purchase a variety of wines through online platforms. However, wine, being a product characterized by highly individual preferences and requiring specialized knowledge, coupled with the explosive growth of online information, has made it challenging for consumers to find suitable wine products online. Therefore, establishing a personalized recommendation system is particularly crucial to enhance consumer experience and increase sales. This study utilizes sales data from a wine e-commerce platform and adopts the well-known Matrix Factorization model along with Bayesian Personalized Ranking algorithm to establish a wine recommendation system for the platform. However, despite the model’s ability to achieve good predictive performance, the recommendations it generates are often difficult to interpret, leading to potential mistrust from companies or users. To address this issue, this study introduces Partial Least Squares Regression (PLSR) to analyze the recommendation logic of the model from the perspective of wine label attributes. Additionally, Hierarchical Clustering is utilized to group the latent factors of the wines generated by the model, with a focus on identifying key wine attributes that contribute to cluster formation. Through these analyses, not only are reasonable explanations provided for the recommendation results, but also trust in the recommendations from the e-commerce platform is enhanced, alongside gaining deeper insights into the consumption behavior of platform users, thus providing valuable references and bases for recommendation strategies.