The goal of transfer learning is to transfer information learned from the source domain to the target domain. A transfer learning method, Optlearn, is proposed for the case of the target domain being a sub-manifold of the source domain. The source domain is weighted to make the weighted source domain and the target domain as similar as possible. The optimal transport theory is employed to minimize the difference between the weighted source domain and the target domain. Furthermore, the dual-Sinkhorn divergence is improved to suit the sub-manifold. Meanwhile, a fast computing algorithm is proposed for Optlearn. The proposed algorithm is tested on the task of pedestrian counting. Experimental results show that Optlearn obtains good counting accuracy as well as avoids the high cost of labeling data for each fixed camera.

For the low-rank matrix and tensor minimization problem, the optimal solution of convex function can be obtained easily, and the better low-rank solution can be obtained from the local minimum of the corresponding nonconvex function. The low-rank tensor recovery problem based on the nonconvex function is studied in this paper. A nonconvex low-rank tensor model based on l_p norm is proposed. In addition, tensor based iteratively reweighted nuclear norm algorithm is proposed to solve the nonconvex low-rank tensor minimization problem. The weighted singular value thresholding problem is solved by the tensor based iteratively reweighted nuclear norm algorithm. The objective function value monotonically decreases and its convergence can be theoretically proved. The recovery performance of the proposed method is demonstrated by comprehensive experiments on both synthetic data and real images.

The acquisition of structural features brings higher complexity to network representation learning. Based on the idea of hierarchy, an effective method is proposed to reduce the complexity of existing network representation learning methods. The network is gradually compressed into a coarse-grained representation space via node neighborhood similarity. And the coarse-grained feature representation is learned by the existing network representation learning methods. Finally, the learned coarse-grained features are gradually refined into the node representation of the original network using the graph convolution network model. Experimental results on several datasets show that the proposed method compresses the network efficiently and quickly, and the running time of the existing algorithms is greatly reduced. For the task of node classification and link prediction, the proposed method can greatly improve the performance of the original algorithm while the granularity level is low.

As the classic hidden Markov model(HMM) loses the sight of confidence of labeled results while building a sequence, a weighted observation hidden Markov model(WOHMM) is proposed. The algorithms in the steps of probability calculation, parameter learning as well as sequence labeling are described in detail. The simulation results on the public datasets show that the parameters obtained by the parameter learning algorithm of WOHMM are closer to the real values than those of HMM, and the performance of sequence labeling algorithm is superior to the state-of-the-art methods.

To narrow the discrepancy between a fuzzy set and its induced shadowed set, a shadowed set model, interval shadowed set, is proposed based on fuzzy entropy. Grounded on the interval shadowed set model, an improved density peak clustering algorithm is proposed to optimize the noise detection strategy of the classical algorithm. To detect the noise, the two-way clustering result of classical algorithm is transformed into three-way clustering result by introducing interval shadowed set model. Finally, comparison experiments on classical artificial datasets and UCI datasets show that the improved algorithm distributes the objects of any dimension and scale more reasonably to the corresponding clusters, and it has good robustness to noise data.

The loss of valuable information may be caused by undersampling, and the class overlapping between the majority class and the minority class may be aggravated by the synthetic minority oversampling technique(SMOTE). A sampling method, Screening_SMOTE, is proposed in this paper, combining safe sample screening based undersampling with SMOTE. Parts of non-informative instances and noise instances in the majority class are identified and discarded by the undersampling method using safe screening rules. Then, the minority class instances generated by SMOTE are added into the screened dataset. The loss of informative information is avoided and the noise instances in the majority class are discarded using safe sample screening based undersampling, relieving the class overlapping. The experimental results show that Screening_SMOTE is an effective method of rebalancing imbalanced datasets, especially for high dimensional imbalanced datasets.

The results of the existing texture recovery methods are not detailed enough for the face features such as wrinkles, beards and pupil colors. Therefore, the texture and illumination preserving 3D face reconstruction based on face normalization is proposed. Firstly, the 2D facial image is normalized to reconstruct the self-occlusion area using illumination information and symmetric texture. Then, the corresponding 3D face texture is obtained by the normalized 2D image according to the 2D-3D point pairs. Finally, combining the face shape reconstruction and texture information, the final 3D face reconstruction results are generated. The experimental results show that the proposed method preserves the texture and the illumination information of the original 2D image effectively, and the reconstructed faces are more natural with more facial details.

Under-sampling is often employed in imbalanced data processing. However, the characteristics of support vector machine(SVM) are seldom taken into account in the existing under-sampling methods,and the problem of losing some key information of the majority class is caused by the sampling in the original space. To solve these problems, a kernel SVM algorithm based on identifying key samples for imbalanced data(IK-KSVM) is proposed in this paper. Firstly, the majority class is divided effectively based on the initial hyperplane. Then, kernel heterogeneous nearest neighbor sampling is conducted on each partition to obtain the key samples of the majority class in the high-dimensional space. Finally, the final SVM classifier is trained by the key samples and the minority class samples. Experiments on several datasets show that IK-KSVM is feasible and effective and its advantages are evident while the imbalance degree of the dataset is higher than 10∶1.