cLP: Linear programming with biological constraints and its application in classification problems

Abstract

Feature selection represents a major challenge in the biomedical data mining problem, and numerous algorithms have been proposed to select an optimal subset of features with the best classification performance. However, the existing algorithms do not take into account the vast amount of biomedical knowledge from the literature and experienced researchers. This work proposes a novel feature selection algorithm, cLP, with the optimized binary classification accuracy. The proposed algorithm incorporates the biomedical knowledge as constraints in the linear programming based optimization model. The experimental data shows that cLP outperforms the other feature selection algorithms, and its constrained version performs similarly well with the unconstrained version. Although theoretically constraints will reduce the classification model performance, our data shows that the constrained cLP sometimes even outperforms the unconstrained version. This suggests that besides the benefit of including biomedical knowledge in the model, the constrained cLP may also achieve better classification performance.