Research On Robust Semi-supervised Learning Algorithm And Its Applications In Biometrics

Semi-supervised learning is an important research field of machine learning, and graph based semi-supervised learning is one of the most promising direction. The core of semi-supervised learning algorithms is graph construction, which has an important impact on the performance of these algorithms. Meanwile the noise and outlier data can also affect the algorithm's performance seriously. Combining with sparse representation theory, correntropy theory and subspace learning theory, a series of robust learning algorithms are proposed in this dissertation, and the effectiveness of these algorithms are demonstrated through extensivly experiments on typical facial datasets. Some of these algorithms are used in two typical biometric identification problems, the microarray tumor recognition and protein secondary structure prediction. These applications further show the improvement of the algorithms, also provide a new solution to biometric identification problems. The main contributions of this dissertation can be concluded as follows.(1) A robust semi-supervised label propagation algorithm based on the non-negative sparse probability graph is proposed and the convergence of the algorithm is proved. Combining with the correntropy theory, the graph construction of the proposed algorithm is implemented through solving an optimization problem representing the training data point as a nonnegative linear combination of the others. The algorithm has strong robustness since it can effectively weaken the interference of noise data. The experimental results on several machine learning datasets show that the algorithm can obtain higher classification accuracy and the robust one has strong robustness.(2) A semi-supervised learning algorithm (Gaussian Laplacian Regularized-Maximum Correntropy Criterion, GLR-MCC) based on the maximum correntropy criterion is proposed and the convergence of it is proved. GLR-MCC has robustness to data noise by replacing the least squares Criterion with maximum correntropy criterion. First the weight of graph is calculated, and second the labels of unlabeled samples are obtained through solving a nonlinear optimization problem. A greedy iteration technique base on half quadratic optimization is adopted to solve the optimization problem. Extensivly experiments on standard facial datasets show that the GLR-MCC algorithm is robust to noise.(3) Aiming at the outliers processing, and using tangent approximation and tangent alignment, a large margin discriminant tangent analysis method as well as an outlier detection algorithm is proposed to learn a decision boundary. A robust subspace is obtained by builting a robust intra-class matrix. Extensive experiments on five face recognition datasets not only demonstrate the effectiveness and efficiency of the proposed algorithm, but also show its practicability of facial recogniton on face datasets especially challenging ones. The research also shows the impact on the performance of the algorithm by choosing threshold parameters.(4) Presently there are not many semi-supervised algorithm be applied to Microarray tumor recognition and protein secondary structure prediction, which are the two typical problems in bioinformatics. The proposed nonnegative sparse representation label propagation algorithm is applied to microarray tumor recognition (as leukaemia, colon and so on) and protein secondary structure prediction in the dissertation. The experimental results show that even with limited labeled data, the algorithm can still obtain satisfactory results. Meanwhile the experiments also show the practicability of the proposed algorithm in biometric identification.