| 摘要: | 近幾年電子商務崛起，國際間貨運往來的需求大增，雖然整體而言，出口貿易額是呈現萎縮的情況，但是國際間貨物交易量體逐年提升。在凡事講求快速的這個時代，貨物透過空運遞送的量體相對來說數量還是非常龐大，如何快速的將空運出口貨物迅速且正確的作業完畢，考驗著每一家倉儲業的能力。過去的作業習慣皆依據人的經驗將無法進入自動化倉儲設備的貨物擺放到適當的位置，但由於人員流動且作業經驗無法完全的傳承，導致出口貨物作業時將花費大量的人力與時間去尋找貨物。本研究的目的包括利用資料探勘監督式學習技術挖掘出空運出口貨物存倉時間長短的預測模型、建立空運出口貨物存倉時間的模型，並進一步比較單一分類技術與多重分類技術之差異及藉由本次研究，提供個案公司在預測空運出口貨物存倉時間的相關產業做為參考。
本研究在實驗流程上採用Weka資料探勘軟體，並進行不同分類技術的實驗，本研究在單一分類技術分別採用決策樹推估模式、支援向量機推估模式、類神經網路推估模式、最鄰近演算法等四種單一分類技術，並搭配多重分類技術中的Bagging、AdaBoost加以驗證，以試圖獲得最佳空運出口貨物存倉時間預測模型。
經過實驗結果得知以2015年的訓練資料集而言，在單一分類技術中以最鄰近演算法表現最佳，在多重分類技術中分別以Bagging的最鄰近演算法、AdaBoost的J48表現最佳，透過Weka的實驗結果，正確率(Correctly Classified Instances)與接收者操作特徵曲線(ROC)普遍值達到0.74、0.8左右，具有較佳參考意義。因此，本研究建議個案公司未來在進行空運出口貨物存倉時間預測時，可以優先採用單一分類技術中的最鄰近演算法，並搭配多重分類技術中Bagging的最鄰近演算法、AdaBoost的J48，以進行空運出口貨物存倉時間預測分析。
;In recent years, the rise of e-commerce, the international demand for freight traffic increases, although the whole, the export trade volume is shrinking, the international trade volume of goods is increasing year by year. In this era of rapid demand, the quantity of goods delivered by air is relatively large, and how quickly the air cargo will be transported quickly and correctly, and the capacity of each warehousing industry is tested. Since past operating habits are based on human experience, the goods are not be able to enter the automated storage equipment and be placed in the appropriate location. Moreover, because of the flow of personnel and operating experience can not be fully inherited, export operations require a lot of manpower and time to find goods. The purpose of this study includes the use of data exploration and supervised machine learning technology to excavate the short time of the export of goods inventory forecast model, the establishment of air cargo export time warehouse model. In particular, single and multiple classification techniques are compared in order to find the optimal model and provide relevant companies with a reference to the industry concerned in forecasting the time of export of goods by air.
In this study, different classification techniques were constructed by the Weka data mining software. Particularly, the decision tree (J48), support vector machine, neural network, and nearest neighbor were used for the single classification techniques. On the other hand, the bagging and AdaBoost methods are employed to construct the multiple classifiers for comparisons.
Experimental results show that in the case of 2015 training data set, the best single classifier is the nearest neighbor algorithm whereas the multiple classifiers are the nearest neighbor algorithm by bagging and decision tree by AdaBoost. More specifically, the receiver operating characteristic curves (ROC) of these classifiers generally reach 0.74, 0.8 or so, with a good reference. Therefore, this study suggests that for the future of the case company in the air cargo export time forecast, they can give the priority to employ a single classification technology based on the nearest neighbor algorithm and multiple nearest neighbor classifiers by bagging the and multiple J48 classifies by AdaBoost to carry out air cargo export time forecasting analysis. |
