Evolutionary Cost-Sensitive Discriminative Learning With Application to Vision and Olfaction

In the design of machine learning models, one often assumes the same loss, which, however, may not hold in cost-sensitive learning scenarios. In a face-recognition-based access control system, misclassifying a stranger as a house owner and allowing entry may result in a more serious financial loss than misclassifying a house owner as a stranger and not allowing entry. That is, different types of recognition mistakes may lead to different losses, and therefore should be treated carefully. It is expected that a cost-sensitive learning mechanism can reduce the total loss when given a cost matrix that quantifies how severe one type of mistake is against another one. However, in many realistic applications, the cost matrix is unknown and unclear to users. Motivated by these concerns, in this paper, we propose an evolutionary cost-sensitive discriminative learning (ECSDL) method, with the following merits: 1) it addresses the definition of cost matrix in cost-sensitive learning without human intervention; 2) an evolutionary backtracking search algorithm is derived for the NP-hard cost matrix optimization; and 3) a cost-sensitive discriminative subspace is found, where the between-class separability and within-class compactness are well achieved, such that recognition becomes easier. Experiments in a variety of cost-sensitive vision and olfaction classification tasks demonstrate the efficiency and effectiveness of the proposed ECSDL approach.