Using Goodness-of-fit Techniques to Select Virtual Samples to Improve Small Data of Cost Prediction

• Main Authors: Wei-LinLiao, 廖洧琳
• Other Authors: Der-Chiang Li
• Format: Others
• Language: en_US
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/64579382380698443165

碩士 === 國立成功大學 === 工業與資訊管理學系碩士在職專班 === 104 === After the transition of consumer demand in the market, in order to create the technology and develop the finance, corporations adopt industrial transformation strategy to make an alteration of the corporation structure and scale and improve the competition and satisfy customer needs. Under stress of time and cost, the cost prediction in the product development process is often being conducted without acquiring sufficient sample data. The method of constructing prediction model to perform cost prediction analysis from limited data in the competitive and qualitative environment has become an important goal for corporation management. For the issue of insufficient sample data in the practice and application, the matter of how to develop the meaningful information from few samples is the research topic of small data set learning. Many studies noted that in order to address the predict issues of production system and break through the learning efficiency when data is inadequate, small data research increase the sample amount by virtual sample generation method to increase the sample data through small data sets and enhance the accuracy of prediction learning. The study adopted methods of two parameters Weibull distribution and Maximal P-value approach to generate virtual sample to increase data. The confounding factors were filtered and screened by Kolmogorov–Smirnov test and imported the training sample data into the prediction learning instruction to compare the accuracy of prediction in different methods. The study adopted small data sets of automotive printed circuit board in Taiwan as experiment case. The result indicated the method conducted in the study had a significant improvement on the prediction error than one in original data, and it received better effect of prediction comparing to maximum p-value.