- 學位論文
應用機器學習進行連鎖速食餐飲業來客數與銷售之預測
Applying machine learning to predict the number of customers and sales for a fast-food chain restaurant
摘要
隨著零售業的數據化轉型，準確預測來客人數、銷售額以及商品銷售量成為業界面臨的重大挑戰。本研究旨在探討和比較三種機器學習模型： Seasonal Autoregressive Integrated Moving Average Exogenous (SARIMAX) 、 eXtreme Gradient Boosting (XGBoost) 和 Long Short-Term Memory (LSTM) 模型在使用零售數據對於需求預測上的表現。通過對一家連鎖速食企業過去半年的日銷售數據進行分析。 本研究對這三種模型進行了模型性能比較， 預測短期的來客人數、銷售額及商品銷售量。研究方法包括數據預處理、特徵選擇、模型訓練與優化以及預測性能評估。首先，對原始數據集進行清洗和預處理，以滿足模型訓練的需求。其次，利用相關性分析和特徵重要性評估來選擇最具預測價值的特徵。然後，對SARIMAX、XGBoost 和 LSTM 模型進行參數調優，以達到最佳預測效能。最後，通過比較各模型的預測結果，包括 R2、 MAPE、 MAE 和 RMSE，來評估和分析在不同預測任務上的表現。本研究的結果表明， 在來客人數為 SARIMAX 模型表現最佳，銷售額為 XGBoost 模型最佳，商品的銷售數量表現最佳的皆是SARIMAX 模型。 透過來客人數預測結果我們可以進行員工人力排班，銷售額預測結果可以對公司的規劃進行布局，商品銷售數量預測則是可以對於成本進行有效的控管。
並列摘要
As the retail industry undergoes digital transformation, accurately predicting foot traffic, sales revenue, and product sales volume has become a significant challenge. This study aims to explore and compare three machine learning models: Seasonal Autoregressive Integrated Moving Average Exogenous (SARIMAX), eXtreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) models in their performance on demand forecasting using retail data. Analysis was conducted on the daily sales data of a chain fast-food enterprise over the past six months. The study compares the performance of these three models in predicting short-term foot traffic, sales revenue, and product sales volume. Research methods include data preprocessing, feature selection, model training and optimization, and predictive performance evaluation. Firstly, the original dataset was cleaned and preprocessed to meet the requirements for model training. Secondly, the most predictive features were selected through correlation analysis and feature importance evaluation. Then, parameters of the SARIMAX, XGBoost, and LSTM models were tuned to achieve optimal predictive performance. Finally, the predictive results of each model were compared, including R2, MAPE, MAE, and RMSE, to assess and analyze their performance in different forecasting tasks. The results of this study indicate that SARIMAX performs best in predicting foot traffic, XGBoost performs best in predicting sales revenue, and SARIMAX also performs best in predicting product sales volume. Through the prediction of foot traffic, we can schedule employee shifts, sales revenue prediction can inform company planning, and product sales volume prediction can effectively control costs.