Analysis and simulation of a Feature Importance Based Structural Correspondence Learning algorithm

In traditional text classification, training and testing text are assumed to be Independent and identically-distributed. With emerging product reviews on E-commerce websites, text classification applied to these domains no longer obeys the IID assumption. At the same time, many transfer learning algorithms are proposed to solve this problem. This paper proposes a framework focusing on feature importance study, which a representative transfer learning algorithm is embedded into. The experimental results show that this frame can significantly improve the transfer learning performance of the embedded algorithm, and feature importance study has a potentially important role in transfer learning. By studying the impact of FIB-SCL between the A-Distance, FIB-SCL was found to reduce the A-Distance between the source and target text.