應用資料探勘技術於資源配置預測之研究-以某電腦代工支援單位為例

范睿昀(Jui-yun Fan)

指導教授 ： 蔡志豐

國立中央大學/管理學院/資訊管理學系碩士在職專班/碩士(2015年)

摘要

隨著商業模式快速的改變，國內外電腦代工產業面臨產業的快速脈動、少量多樣的市場需求，紛紛尋求如何提升產品協同設計流程效能的方法。因應產品「少量」、「多樣」的生產模式策略，支援單位即將邁入多樣少量日趨複雜產品設計，其資源配置與協同合作流程監控的提升日趨重要。
學術界對於政府單位的資源配置提供很多的預測理論與技術，但在電腦代工領域預測卻相對較少探討。本研究以個案公司龐大的電子料件資料庫，進行「單一與多重分類器」資源配置預測效能比較。試圖找出符合個案公司需求之最適工具並期許本研究結果能提供學術界參考。
本研究在實驗流程上，採用Weka資料探勘工具，進行不同分類器實驗，為能找出個案公司資料中具備影響力的維度，主要是透過基因演算法 (GainRatio AttributeEval) 和資訊增益 (Information Gain) 去執行資料前處理，以試圖獲得最佳資源配置預測模型。
經過實驗結果得知，無論是單一分類器或多重分類器，ROC值的整體表現皆高於0.9並以分類器之CART的正確率較高，其特徵屬性選取 (Feature Selection) 對於資料集的預測結果影響都不大，因此，本研究建議，個案公司在處理資源配置的預測時，可以優先採用分類器之分類決策樹推估模型 (CART)。

關鍵字

資料探勘 ； 資源配置 ； 特徵屬性選取 ； 單一與多重分類器

並列摘要

With rapid changes of business systems, both domestic and foreign industrial manufacturers face the problem of fast flow and various market demands. As a result, they are looking for the solution which could increase the effectiveness of cooperative product design. Due to the two production strategies, which are “low quantity” and “various”, the support department will design a variety of products with a small quantity. Also, its resource allocation and cooperative process supervision will become more and more important.
Academics provide numerous predictive theories and techniques for the resource allocation of government department. However, there are less predictions for the field of computer industrial manufacturers. This research uses a lot of electronic resources from a case company to create difference single and multiple classifiers for predicable resource allocation. This is to identify the most suitable model for the demand of the company and we expect that the result of the research could be referred by academics.
For the experiments, the research uses the Weka data mining tool to conduct the experiments for different allocators. In order to find out the most effective dimension of the company, the research use the GA-based heuristic algorithm and information gain algorithm to pre-process the data first. Therefore, we could obtain the best predictive model of resource allocation.
From the experimental results, we observe that no matter single or multiple classifiers are used, the general performance of R.O.C. are all higher than 0.9, and the bagging based multiple classifiers based on CART got a higher accuracy, but performing feature selection has a little influence on the prediction of data collection. Therefore, the research recommends that when making a prediction of resource allocation, the company could use the classifier for the allocation strategy based on CART.

並列關鍵字

Data Mining. ； Resource-Base View ； Feature Selection ； A Single and Multiple Classifier

參考文獻

3.林明潔，董子毅 (2008)，「危險評估中ROC曲線在預測 2×2 表上與敏感度及特異度之關係」，亞洲家庭暴力與性侵害期刊，第四卷第二期， 64 -74頁。

18.楊正三，葉明龍，莊麗月，陳禹融，楊正宏(2008)，「利用資訊增益與瀰集演算法於基因微陣列之特徵選取與分類問題」，資訊科技國際期刊，第二卷，第十期，第50-62頁。

11.郭光明，陳瑞甫，黃興進，劉忠峰 (2011)，「影響醫學影像儲傳系統(PACS)績效關鍵因素之研究—資源基礎觀點」，資訊管理學報，第十八卷，第三期，50-625頁。

25.簡禎富、林國勝 (2006)，「建構cDNA生物晶片之二元資料挖礦模式及其實證研究」，資訊管理學報，第十三卷，第四期， 133-159頁。

2.林清河 (2007)，物料管理，台北，華泰書局。