Prediction performance improvement for highly imbalanced monitoring data

In engineering applications, we often face highly imbalanced data problems where majority of the data are from a condition and small minority are from others. Directly learning classifier on such problems would be prone to a biased classification performance by the majority class, so resulting in poor predication on the minority class. This paper proposes a method for balancing training data, which over-samples the minority class. The method uses between-class and within-class information to decide the vicinity space of an example. It generates synthetic examples along orthogonal directions in the vicinity, so it ensures the generated synthetic examples well represent the entire vicinity space and be more similar to minority class than majority class. The method is easy to use, as it involves no parameter setting. A real world problem of semiconductor manufacturing line monitoring and process control data is used to demonstrate that classification performance can be significantly improved through learning on balanced data by the proposed method.