In semi-supervised classification, labels smoothness and cluster assumption are the key point of many successful methods. In graph-based semi-supervised classification, graph representations of the data are quite important. Different graph representations can affect the classification results greatly. Considering the two assumptions and graph representations, we propose a novel method to build a better graph for semi-supervised classification. The graph in our method is called m-step Markov random walk graph (mMRW graph). The smoothness of this graph can be controlled by a parameter m. We believe that a relatively much smoother graph will benefit transductive learning. We also discuss some benefits brought by our smooth graphs. A cluster cohesion based parameter learning method can be efficiently applied to find a proper m. Experiments on artificial and real world dataset indicate that our method has a superior classification accuracy over several state-of-the-art methods.
[1]
Bernhard Schölkopf,et al.
Cluster Kernels for Semi-Supervised Learning
,
2002,
NIPS.
[2]
Ronald Rosenfeld,et al.
Semi-supervised learning with graphs
,
2005
.
[3]
Alexander Zien,et al.
Semi-Supervised Classification by Low Density Separation
,
2005,
AISTATS.
[4]
Alexander Zien,et al.
Semi-Supervised Learning
,
2006
.
[5]
Joseph Rudnick,et al.
Elements of the random walk
,
2004
.
[6]
W. D. Ray,et al.
Stochastic Models: An Algorithmic Approach
,
1995
.
[7]
Tommi S. Jaakkola,et al.
Partially labeled classification with Markov random walks
,
2001,
NIPS.
[8]
Zoubin Ghahramani,et al.
Combining active learning and semi-supervised learning using Gaussian fields and harmonic functions
,
2003,
ICML 2003.
[9]
Bernhard Schölkopf,et al.
Learning with Local and Global Consistency
,
2003,
NIPS.