- 學位論文
單一與多專家銷售預測模型比較
A Comparison of Single and Multipal Sales Prediction Models
摘要
現今預測的方法大致可分為兩方向，第一種為傳統的統計分析，如迴歸分析、相關分析、區別分析、Logit和Probit模型等，第二種為計算智慧的資料探勘工具，如決策樹、類神經網路、支援向量機、灰色預測、模糊理論等。現今眾多的流行預測方法中，不同的領域可能衍生不同的屬性，故多數學者在眾多的預測方法與工具中，皆是針對特定的資料集，選定攸關的資料屬性後採用單一種的預測工具與其他工具進行比較，較少學者透過多種的預測方法及預測模型進行統計評比與彙整分析。故本研究使用多種單一計算智慧預測工具如：支援向量機、類神經網路與傳統預測方法，如：移動平均法、最小平方法、保序迴歸法、簡單線性迴歸，並同時使用單一計算智慧預測工具結合Bagging和Stacking 集成學習方式，做為本研究之中繼分類器以進行研究實驗，再以平均絕對誤差百分比做為本研究各預測器之評估標準，評選出各模型中最適合本研究資料集的預測器，以提供學者爾後在採用方法時有不同的依據。本研究實驗結果得知傳統方中以最小平方法最好，而計算智慧預測工具中單一計算智慧預測模型以BPN預測器，較適合本研究的資料集，Bagging集成預測模型中以BaggingSLR預測器，較適合本研究的資料集情境，Stacking集成預測模型則以Stacking結合Bagging BPN預測器，為最適合本研究資料集的預測器，彙整上述四類模型實驗之結果，Stacking結合Bagging BPN在此16種預測器中表現最好。
並列摘要
Generally speaking, there are two groups of methods in the field of forecasting, i.e., traditional statistic analysis and artificial intelligence techniques. The first includes the regression analysis, correlation analysis, discriminate analysis, logit model, probit model, etc. The second includes decision tree, neural network, support vector machine, gray prediction, fuzzy theory, etc. Such methods may need different attributes for their specific application. Therefore, in literature, most forecasting tasks focus on specific attributes in their specific problem domains. In particular, they only use one or two forecasting methods to compare with one or two other methods after relevant attributes have been determined. The scale of such comparison is not wide enough. Thus, this thesis firstly uses multiple forecasting tools in artificial intelligence field, such as support vector machine (SVM) and neural network (NN). Secondly this thesis uses traditional statistical forecasting methods, such as simple linear regression (SLR), moving average (MA), ordinary least squares (OLS), and isotonic regression. Thirdly this thesis uses hybrid approaches by integrating artificial intelligence approaches with traditional statistical forecasting approaches. Finally, we also combine bagging and stacking ensemble learning techniques to get an improvement for our problem domain in the thesis. The standard of evaluation used in the thesis is the mean absolute percentage error (MAPE) that evaluates the most suitable forecaster from those (i.e., 16) forecasters in our specific problem domain which may provide more complete concept when similar forecasting task is performed in the future. The results in the thesis show OLS is the best model in the traditional statistical group, the NN forecaster is the best model in the artificial intelligence group, the SLR bagging forecaster is the best model in bagging group and the method which integrates NN bagging with stacking model is the best one in sixteen model.