一個採用選擇性集成的零售商品預測模型

A SELECTIVE ENSEMBLE OF SALES PREDICTION MODELS

陳勁宏(Jing-Hong Chen)

指導教授 ： 洪智力

中原大學/商學院/資訊管理學系/碩士(2008年)

https://doi.org/10.6840/cycu200900485

摘要

時間數列(Time Series)的價值一直被學術界與實務界所重視，然而現實世界的時間數列是充滿噪音(Noise)與非穩定性(Nonstationary)之問題。本研究使用傳統的線性迴歸預測器(Linear Regression)和人工智慧(Artificial Intelligence, AI)預測工具，如：支援向量機(Support Vector Machine ,SVM)、類神經網路(Back Propagation Neural Network)，透過三種集成策略，分別是(1)中位數選擇法之選擇性集成、(2)時間落差集成預測法、(3)時間落差選擇性集成預測法，以解決三個問題(A)單一預測器易受噪音影響的問題、(B)訓練樣本需侷限於近期資料，因此訓練樣本不足之問題、(C)解決移動平均(MA)所產生的時間落差之問題。實驗結果顯示第一種策略「中位數選擇法之選擇性集成」能改善「缺點A-單一預測器易受噪音影響的問題」，透過具有時間差異的三個單獨預測模型(預測器i+1、預測器i+2、預測器i+3)來集成，進而有效降低C公司的50個商品小分類之預測的平均絕對誤差百分比(Mean Absolute Percentage Error, MAPE)，整體表現上，中位數選擇法(SMO-reg)能有效降低單獨預測器模型(即預測器i+1、預測器i+2、預測器i+3)分別為4.50%、14.99%、8.97%MAPE的誤差率；另外，中位數選擇法(BPN)能有效降低單獨預測器模型(即預測器i+1、預測器i+2、預測器i+3)分別為0.38%、8.39%、7.35%MAPE的誤差率。另外，第二、第三種集成策略能改善「缺點B-訓練樣本需侷限於近期資料，因此訓練樣本不足之問題」、「缺點C-解決移動平均(MA)所產生的時間落差之問題」兩個缺點，整體表現而言，具有日期作為輸入變數的BPN與線性迴歸(Linear Regression)劣於未輸入日期變數的BPN與線性迴歸，而預測表現最好的是未輸入日期變數的BPN並透過第三策略所集成的預測模型，能提供28.19%MAPE的誤差率，降低了移動平均法(Moving Average)3.52%MAPE的誤差率，並在分類預測漲或跌的評估準則下，提供77.78%預測準確率(classification accuracy, CA)，有效改善移動平均法為落後指標之缺點。

關鍵字

選擇性集成學習 ； 支援向量機 ； 類神經網路 ； 時間數列

並列摘要

Although the task of time series is full of noise and non-stationary, the value of its proper exploitation has attracted both researchers and practitioners. This thesis uses tools from the field of artificial intelligence (AI) such as the support vector machine (SVM) and the back propagation neural network (BPN) in order to predict the non-stationary movement of time series. More specifically, three ensemble strategies, i.e. the median based selective ensemble, the time-lag based ensemble, and the time-lag based selective ensemble are used. These three ensembles are designed to deal with the three problems, i.e. the low accuracy predicted by a single classifier due to the noise of data, not enough training samples as only data samples located near to the target sample are useful, the time lag problem of the traditional moving average (MA) approach. The first ensemble strategy handles the first problem successfully. The second and third ensemble strategies overcome the other problems. According to the experimental results from 50 small categories of products of the C company, the proposed ensemble strategies are able to deal with such three problems and therefore improve the prediction performance evaluated by the mean absolute percentage error (MAPE).

並列關鍵字

Back Propagation Neural Network ； Time Series ； Support Vector Machine ； Selective Ensemble.

參考文獻

▪ 李宏志、邱至中，“自我迴歸整合移動平均- 倒傳遞類神經網路與基因演算法在短期匯率預測績效之比較”，財務金融學會年會暨學術論文研討會，2005。

▪ Bates, J.M., C.W.J. Granger, “The combination of forecasts,” Operational Research Quarterly, 20, pp.451-468, 1969.

▪ Box, G.P., and Jenkins, G.M., “Time Series Analysis: Forecasting and Control,” Holden-day Inc., 1976.

▪ Breiman, L., “Bagging predictors”, Machine Learning, Vol. 24, No. 2, pp.123-140, 1996.

▪ Chakraborty, K., Mehrotra, K., and Mohan, C.K., “Forecasting the behavior of multivariate time series using neural networks,” Neural Networks, 5, pp.961-970, 1992.