A convex multi-class model via distance metric learning based class-to-instance confidence

Highlights：

• A novel convex multi-class model via DML-based C2I confidence (DM-C2IC) is proposed. To further boost the classification performance, DM-C2IC is extended to the kernel space.

• DM-C2IC is a non-kNN method for multi-class DML problems and only needs to store the DML-based C2I confidence model of which the size is Kd. More importantly, DM-C2IC can avoid the influence of the nearest neighbor number k.

• A generalized block coordinate descent optimization approach is developed for DM-C2IC. The optimization approach can train and update the DML-based C2I confidence model and the specific Mahalanobis distance metric alternately.

• Extensive experiments are conducted to evaluate the performance of the DM-C2IC model. The better results indicate that DM-C2IC achieves explicitly accuracy improvements than the existing kNN DML methods.