A Multikernel-Like Learning Algorithm Based on Data Probability Distribution

Figure 4

Two-class problem experiments: in (a, b, c, and d), (a) and (b) report the error rates () of MKDPD and MR on unlabeled and test samples of the USPS set, respectively, and (a) and (b) report the error rates () of MKDPD and MR on unlabeled and test samples of the MNIST set, respectively, where the -axis is the 45 binary classification problems; in (e, f, g, and h), we report the scatter plots of error rate on unlabeled samples versus test samples. (e) and (g) are the scatter plots of error rates of MKDPD on USPS and MNIST, respectively, and (f) and (h) are the scatter plots of error rates of MR on USPS and MNIST, respectively.

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)