| dc.description.abstract | Multiple function devices are a type of office machines which combines E-mail, fax,copy, printing, and scanning functions. It was designed to provide users with easy and promptoperation and usage. In the literature of data mining applications, very few focus on B2B selling forecast in Taiwan. Moreover, there is no a comparative study for the applicability of data mining techniques to different types of forecasting results, which are continuous and discrete prediction outputs. Therefore, in this thesis the research objective is to compare different supervised learning techniques for the sale forecast of multiple function devices. The contributions of this thesis are able to provide some guidelines for the case company to conduct sales forecast and can give academics a reference on B2B industry.
In the experiments, the attributes relate to sales from historical data are collected, and the data completeness in each attribute is also taken into account. Next, the historical selling quantity (i.e. continuous values) is used as the prediction output. In addition, the selling quantity is further divided into 3 classes by normal distribution for comparison. On the other hand, in order to find out the effect of performing feature selection on the forecasting result,PCA (principle components analysis) is used to select more representative attributes from the original data set. For model construction, different single and multiple classification techniques are compared.
The experimental results show that performing feature selection does not significantly affect the final prediction results no matter for continuous or discrete prediction output. For continuous prediction without PCA, the support vector machine (SVM) performs the best in terms of MAE (Mean Absolute Error). For discrete prediction without PCA, the SVM outperforms the other models in terms of prediction accuracy. | en_US |