Recipe Generation from Small Samples: Incorporating Weighted Kernel Regression with Artificial Samples

The cost of the experimental setup during the assembly process development of a chipset, particularly the under-fill process, can often result in insufficient data samples. In INTEL Malaysia, for example, the historical chipset data from an under-fill process consist of only a few samples. As a result, existing machine learning algorithms for predictive modeling cannot be applied in this setting. Despite this challenge, the use of data-driven decisions remains critical for further optimization of this engineering process. In this study, a weighted kernel regression with artificial samples (WKRAS) is introduced to improve the predictive modeling in a setting with limited data samples. In the proposed framework, the original weighted kernel regression (WKR) is strengthened by incorporating artificial samples to fill the information gaps between available training samples. The artificial samples generation is based on the dependency measurement between every independent variable and dependent variable with subject to the calculated correlation coefficients. Even though only four samples are used during the training stage of the setup experiment, the proposed technique is able to provide an accurate prediction within the engineer’s requirements as compared with other existing predictive modeling systems, including the WKR and the artificial neural networks with back-propagation algorithm (ANNBP).