近年來,隨著新零售的崛起,加上疫情的肆虐,不少行業被迫進行數位轉型,結合人工智慧、大數據等工具,使得新零售技術逐漸成熟,不但整合同步線上、線下庫存資訊,更減少滯銷、缺貨的損失,加上方便的物流,能夠因應線上需求,短時間進行線下補貨,提供顧客更好的消費體驗,因此掌握新零售優勢,將成為各行業脫穎而出的關鍵。本研究以Y公司所研發之智慧販賣機作為研究對象,針對54個設立智慧販賣機之捷運站進行分析,以2019年12月至2020年10月官方網站線上庫存數據為依據,透過捷運站智慧販賣機的角度,延伸計算每月之LRFMP及顧客活躍性指標(Customer Activity Index, CAI)並置於資料群內。接著,本研究針對銷售點機台之設址位置(即捷運站別)進行分群與各站智慧販賣機之銷售預測,前者使用K-means、階層集群分析法、二階段集群分析法個別將銷售點機台分群成不同的群體並進行三者之分群比較,除將各群潛在價值高低程度排序之外,也分別針對各群給予不同的行銷建議,期望能提供Y公司行銷上的幫助;後者使用隨機森林演算法與梯度提升樹演算法建置銷售預測模型,隨後比較兩種模型之準確率高低並給予適當之模型建議,期望能減少供不應求的損失與供過於求的浪費,以提升整體獲益。結果顯示,(1)三種分群方法均有共通分群;(2)定義共通分群並針對各群潛在價值高低給予不同行銷建議;(3)兩種銷售預測模型皆有良好之預測能力;(4)於限制條件下,梯度提升樹較隨機森林適合預測數值。
With the rise of online retail and the outbreak of epidemic diseases, many enterprises have been forced to undergo digital transformation. Meanwhile, cutting-edge retail technology, artificial intelligence, and big data have reduced inventory overstocks and out-of-stocks, allowing short-term offline replenishments in response to online demand. There is no doubt that utilizing the benefits of new retail will be vital to the success of a wide range of industries. This study examines the sales and inventories of vending machines at MRT stations, using data collected from the official website. The Customer Activity Index (CAI) and LRFMPs are calculated monthly and placed in data groups for analysis. Specifically, our analysis covers the MRT station cluster and each station's smart vending machine sales forecast. First, we apply K-means, Hierarchical Cluster Analysis, and Two-stage Cluster Analysis for smart vending machines clustering and compare their results. Then, we apply the Random Forest algorithm and the Gradient Boosted Trees algorithm to build the sales forecast models, compare the accuracy of the two models and give appropriate model suggestions. The results show that (1) all grouping methods have common groupings. (2) Define common clusters and give different marketing suggestions for each group's potential value; (3) Both sales forecasting models have good forecasting ability; (4) Under restricted conditions, the Gradient Boosted Trees model is more suitable for predicting values than Random Forest model.