||March 12, 2007
||Prof. Shin Ishii, Graduate School of Information Sciece, Nara Institute of Science and Technology
||Machine learning approaches to gene expression analyses
|| Analyses of gene expression profiling data are challenging, because of the relatively
small amount of data in comparion to the underlying high-dimensionality.
In this talk, I first present a missing value estimation method based on Bayesian
principal component analysis (PCA), which assumes that the gene expression data are
generated by adding high-dimensional noise to low-dimensional factors.
Next, I present a correlation-based clustering method whose component model is a
constrained version of probabilistic PCA. Bayesian estimation provides a statistically
consistent estimation of the model and parameters from actual data. Finally, I present
a multi-class classification method by integrating binary classifiers based on the
analogy from information transmission theory. The results when applied to cancer
classification problems from gene expression profiling are shown.