Improving the accuracy of multiclass classification in machine learning: A case study in a cell signaling dataset

Abstract

It is important to make sense of the data within its context to propose a useful model to solve a problem. This domain knowledge includes information not contained in the data, but that will help us understand the data to be fed into a machine-learning algorithm and guide us on what features might help our model. Nevertheless, domain knowledge may become insufficient as the input variables increase, forcing the need to try automated feature selection techniques. In this study, we investigate whether the joint use of 1) feature selection techniques, such as Chi-square, Tree-based Feature Selection, Pearson’s Correlation, LASSO, Low Variance, and Recursive Feature Elimination, 2) outlier detection methods such as Isolation-Forest, and 3) Cross-Validation techniques lead to improving the accuracy in multiclass classification in machine learning. Specifically, we address the classification of patterns representing the activation state of cell signaling components into classes that symbolize the different cellular processes triggered in cancer cells. The results presented in this work have shown an accuracy increase with up to 80% fewer input features by only using 3 out of the 16 original descriptors.