運用資料探勘輔助商品分類之需求預測方法

Abstract

商品分類為所有與管理商品相關資訊活動的核心，每個公司都會為商品分類，用於銷售管理、採購管理、存貨管理等方面。需求預測是需求管理中最重要的功能，過去許多研究提出改進傳統時間序列趨勢預測方法的準確度，其中之一是合併不同商品的銷售記錄，降低資料的變異度，以合併後的資料進行預測再用適當的比率分配給各個商品。合併商品銷售記錄的依據便是商品分類。以往管理者以質性觀點所建立的商品分類架構並非完全適用於需求預測，貿然將銷售發展趨勢差異太大的商品歸為同一類，會導致類別商品的發展趨勢扭曲或模糊。本研究希望以量化觀點輔助調整商品分類架構使其適合需求預測所用。 針對此問題，本研究根據資料探勘分類方法中的距離基礎方法，定義商品之間銷售發展趨勢的相似程度，進而提出一兩階段最佳化目標模式：首先在固定分群數目下最小化群集內樣本間距離總平均；然後比較不同分群數目時，最大化群集間距離總平均。 本研究的最佳化目標函式並非線性，且解集合的型態近似於整數規劃，必須在每個整數點上搜尋，可行解區域大小隨著分類的商品數目呈現指數速度成長，因此無法利用有限資源求出最佳解。本研究提出一啟發式演算法，使上述問題在可接受的時間內找到一趨近最佳解的分類結果。 本研究啟發性演算法主要流程為：在前置作業中，以時間序列分析進行分類所需之資料轉換，然後根據前一步驟的分析結果建構分類階層，並以基因演算法為基礎搜尋最適分類結果。新的分類架構匯入一需求預測學習系統進行預測準確度評估。 最後，本研究實作出此分類架構建立系統，以兩個實際案例進行驗證本研究所提出之方法確實可行且具有效率。經過實驗之後發現，數百個商品若擁有長期的銷售歷史，且具有明顯的長期趨勢與季節性波動，經過本研究所提之方法分析可以有效判別商品之間的異同，並加以群集，提升預測準確度。本研究的適用商品不限產業，也足以應用於供應鏈管理其他功能。

Product classification is the core of every information activity related to product management. Almost all companies classify their products according to some attributes for different management purposes such as sales, procurement, and inventory control. Within these business functions, demand management is the leading pulling force while demand forecasting is the most critical function of demand management. Previous studies have suggested many ways to improve the accuracy of prediction using traditional time-series analysis with trend, and one of notable techniques is aggregating sales records of individual product. The purpose of sales aggregation is to reduce the data variation, which can then result in a better sales forecast. Product classification can be used as the scheme for deciding which items should be combined into one product class. Most companies cluster or group their products based on qualitative features such as brand, color, package, etc, even though for different purposes. The sales trends might be distorted or become unremarkable if the products are carelessly clustered together. This study aims to cluster products by analyzing their quantitative characteristics, namely sales pattern, and make it more suitable for demand forecasting. This study defines the similarity of sales pattern among various products by adopting distance-based method in data mining, and furthermore develops a two-phase optimization model: starting with a given number of groups, minimizing the average distance within groups, then looking for maximization of average distance among separated groups through incrementing number of groups assigned. Because of the non-linear nature of objective function, integer programming is a popular way to solve the problem. However, when the number of items to be classified increases, the size of feasible solution set grows exponentially as well and makes the problem insolvable due to the time and computing resource it requires. To conquer the difficulty, this study proposes a heuristic algorithm, called Data-Mining Aided Product Classification (DMAPC). DMAPC first analyzes sales records using time-series analysis and transfers them into a number of indexes which can best describe their patterns. Then, DMAPC searches the optimal product grouping result using GA-based heuristic and the extracted indexes from first stage. A demand forecasting learning platform is used in the final stage. In order to show the effectiveness and efficiency, a prototype was constructed and tested to demonstrate the power of DMAPC using complexity and computational analysis.