基于圖的半監(jiān)督機器學習.pdf

1、Over the past few years, semi-supervised learning has gained considerableinterest and success in both theory and practice. Traditional supervised machinelearning algorithms can only make use of labeled data, and reasonab

2、le performanceis often achieved only with a large number of labeled data. However, labeled datais often expensive and time consuming to collect, while unlabeled data is usuallycheaper and easier to obtain. The strength o

3、f semi-supervised learning lies in itsability to utilize a large quantity of unlabeled data to effectively and efficientlyimprove learning performance. Recently, graph-based semi-supervised learning algorithms are be

4、ing inten-sively studied, thanks to its convenient local representation, connection with othermodels like kernel machines. Graph Laplacian is the central quantity of graph-based semi-supervised learning, which plays a r

5、ole in exploring the underlyingmanifold geometry of the data. Using graph Laplacian to form the regularizationproblem and further employing the kernel techniques is a promising approach ofsemi-supervised learning. Th

6、e author first introduce the basic concepts of scmi-supervised Learning. aswell as the utilized tools and theory, such as support vector machines, kernelmethods and regularization theory. The main contributions of th

7、is thesis are mainly presented in chapter 5 andchapter 6. In chapter 5, the author first investigate a class of graph-based semi-supervised learning methods by spectral transformation. Then the formulation ofsemi-supervi

8、sed spectral kernel learning based on maximum margin criterion withspectral decreasing order constraints is formed, and he also maintain that the max-imum margin criterion is a more essential goal of semi-supervised kern

9、el learningthan kernel target alignment by theoretical analysis. By equivalently transform-ing the resulted intractable optimization problem into a quadratically constrainedquadratic programming, the problem can be effic

10、iently solved. Moreover, the au-thor also propose a method to automatically tune the involved trade-off parameter.Furthermore, the author seek another way to learn the spectral coefficients from amore essential view. Due

11、 to the fact that the spectral order constraints are actu-ally not hard requirements but only for the purpose of ensuring the smoothness ofthe score function, the author leaves out those constraints場directly including th

12、esmoothness regularizer into the maximum margin objective, which coincides withthe theory of manifold regularization. Its efficient iterative algorithm is also de-signed next. Experimental results on real-world data sets

13、 have demonstrated thatboth of his proposed spectral learning methods achieve promising results againstother approaches. Motivated by the requirements of many practical problems, in chapter 6 theauthor turns to study

14、 the problem of semi-supervised learning with structuredoutputs, which is a more general topic than the standard semi-supervised learn-ing. By extending the definition of smoothness regularizes to mufti-class setting,he

15、next explore the mufti-class semi-supervised classification. Although the ob-tamed data dependent kernel similar to that of Sindhwani et al., his mufti-classmodel really extend the theory of theirs. Still next, the autho

16、r further generalizethe mufti-class manifold regularization problem to the scenario with structuredoutputs, and the corresponding dual problems are also obtained. From the dualformulations, we can find that the semi-supe

17、rvised learning task finally can beachieved by the supervised structural prediction with a newly defined "data de-pendent joint kernel matrix". This data dependent kernel matrix generalizes thatof Sindhwani et aL to str

18、uctural prediction. Moreover, his proposed inductiveapproach can naturally predict the unseen data points other than the unlabeleddata. Some experiments on text categorization with hierarchies are conducted,and the empir

19、ical results show his approaches actually utilize the structural andmanifold information of the data simultaneously, and finally help us to improvethe prediction performance. As a supplement, the author also proposes the