整合集群技術與機器學習技術建構電腦產品銷售預測模式

由於資訊技術的迅速發展與消費者需求的快速變化，導致如何對電腦產品建構一個有效且精確的銷售預測模型對資訊通路而言是一個重要且關鍵的問題。本研究整合集群技術(Clustering)與機器學習(Machine learning)方法建構集群式預測模式於電腦產品銷售預測。所提方法先利用集群技術將原始資料中具有相似特徵的銷售資料期間分至同一群，之後對待預測資料區間，在找出其與原始資料中最類似的集群後，用這些集群資料以機器學習技術建構預測模式來進行銷售預測。本研究所使用的集群技術包括自我組織映射圖網路(Self-Organizing Map, SOM)、成長型階層自我網路映射圖(Growing Hierarchical Self-Organizing Map, GHSOM)及K-means分群法；機器學習方法則是支援向量迴歸(Support Vector Regression, SVR)及極限學習機(Extreme Learning Machine, ELM)。由於所提之集群式預測模式能針對待預測資料利用集群技術找出最適當的訓練資料以建構預測模式，因此能提高預測之準確度。本研究以台灣某資訊通路之筆記型電腦（NB）、液晶螢幕(LCD)及硬碟（HD）之銷售資料為實證研究對象。研究結果顯示，所提的整合GHSOM+ELM銷售預測模式相較於單一預測模式，在三個資料中均能有較佳的預測績效，顯示將集群工具整合在預測工具上，能有效的降低銷售預測的誤差，提升預測準確度。

關鍵字

銷售預測；集群演算法；機器學習技術；支援向量迴歸；極限學習機

並列摘要

Sales forecasting is an important issue for most enterprise in every industry. It will influence how many products need to be replenished and what price should be made. However, changes in the electronic industry are rapider than others, so forecasts are even more difficult, how to construct a good prediction model for electronic products is a very important thing. In this research, sales forecasting models combining clustering techniques and machine learning approaches are proposed. In the proposed methods, first, the clustering techniques including SOM (Self-Organizing Map), GHSOM (Growing Hierarchical Self-Organizing Map) and K-means method are used to divides the dataset into clusters that assemble the periods with similar characteristics in same group, and then choose the most representative group as the training sample of machine leaning methods including SVR (Support Vector Regression) and ELM (Extreme Learning Machine) to build the forecasting model. The weekly sales data of NB, LCD and HD collected from a present leading computer company in Taiwan are used as illustrative examples. The experimental results showed that the proposed combined clustering-based forecasting scheme –GHSOM+ELM can provide promising forecasting results.