The RF-stacking Model With Under-sampling

In Semiconductor manufacturing,we have multiple levels of testing to test the chip performance.There is huge data from testing,which indicates the chip performance.This paper tries to apply data mining methods in semiconductor industry,to improve the product yield and reduce the manufacturing cost.The semiconductor manufacturing data is unbalanced and it is high-dimensional.Researchers have developed a series of algorithms on unbalanced data or high-dimensional data separately.However,those algorithms can not work well on unbalanced and high-dimensional data like semiconductor manufacturing data.This paper firstly introduces the general solutions of these two problems and proposes the goodness indicators for model and for applied industry.And then this paper describes the actual problem form Semiconductor industry that yield production is a high potential of cost saving project.It analyzes the pros and cons of Random Forest algorithm,and proposes the RF-Stacking model with under-sampling.At last,we get data from the real production line and run model by this RF-Stacking model.Applying our method to the real semiconductor data,we find the proposed method is efficient for unbalanced and high-dimensional scenarios.