深度學習於智慧零售預測模型之研究：以便利商店時效性商品為例

A Deep Learning Approach of Forecasting Model in Smart Retail: A Case Study of Time-Critical Goods at Convenience Store

近年來智慧零售的議題甚囂塵上，許多零售業者整合新興技術亟欲打造新型態的虛實整合零售業，但無論應用或整合了那些資訊技術，其真正用意都是希望能打造一個以大數據重新驅動消費者體驗的新零售模式，以先進科技和資料分析技術改變企業的營運流程與內涵。在便利商店中，時效性商品的銷售佔比重日漸攀升，店鋪管理者雖可透過POS （Point of Sales）系統得知歷史銷售數據和訂購建議進行下單採購，然而時效性商品常因外在環境影響，使得每日的需求有所變動，若能提高商品銷售的預測準確度將可提高營收並降低成本，本研究以便利商店時效性商品為研究對象，嘗試用新的方法來建構更為精準的預測模型。本研究整合多元數據集包括歷史銷售數據、店內促銷活動資訊、外部環境變數等資料，採用深度學習方法建立便利商店時效性商品銷售預測模式，以個案公司過去兩年於12家分店的資料，每家分店各55項的銷售數據作為建立預測模式的資料，結果顯示，使用隨機森林進行特徵變數篩選，以CNN與RNN建立銷售預測模型，並採用MAE和MSE指標進行誤差評比，準確度明顯優於ARIMA、MLF、RBF、SVM之預測結果，顯見深度學習對於預測時效性商品的適用能力確實較為優異。

In recent years, smart retailing has become a hot topic in convenience store industry. Many retailers are integrating in emerging technologies to create a new type of retail business. The intention of those applications or integrations by information technology is to create a new retail model with big data by customer experience. Advanced technology and data analysis technology are adapted to this kind of new retail model to change the operation and process in convenience stores. In convenience stores industry, the percentage of Time-Critical Goods is growing up. Managers can place orders through historical sales data and ordering advice by POS system currently. However, the demand of Time-Critical Goods is impacted due to external environment changing. It will increase profits and decrease unnecessary expenses if the accuracy of forecasting of the sales be improved. The purpose of this study is to use advanced methods to construct more accurate forecasting model for sales of Time-Critical Goods in convenience stores. Mmultivariate datasets including historical sales data, in-store promotion, external environmental variables etc., are integrated in this research. Research team use deep learning methods to establish forecasting model of sales of Time-Critical Goods in convenient store. The data comes from 12 branch stores of case company and each 55 items of sales data is used for models establishing. After using (i) random forests for feature variable screening and (ii) sales forecasting model by CNN and RNN, and (iii) MAE and MSE for error analysis, the accuracy of forecasting is better than the sales forecast model constructed by ARIMA、MLF、RBF and SVM methods. The result showed that deep learning methodology is suitable for forecasting of sales of Time-Critical Goods in convenient store.