Several prominent feature selection and representation techniques, under unsupervised and supervised learning frameworks, will be introduced.  For unsupervised learning, feature selection tools (e.g. PCA, ICA) and cluster discovery (k-means, hierarchical clustering, SOM) play important roles.  For supervised learning, filter  and wrapper techniques represent the prevailing approaches in the current literature.

An emphasis of this talk is on  feature selection.  For this, both individual and group-wise representations have their own pros and cons in measuring the truly relevant information. The individual quantification is simple as each of the M features can be represented by one single value. However, it cannot deal with the inter-feature redundancy, abounding specially in genomic data.  In contrast, the group-wise information can fully address the mutual redundancy, but it is often too complicated to process.  (Note that there are 2M possible groups.)  Between the two extremes, fortunately, there is a convenient compromise: the pairwise kernel - which has a low complexity (M2 pairs) and yet reveals the critical information regarding the m inter-feature redundancy. Indeed, it has already found many useful - and successful - genomic applications.  (For example, in genomic sequencing, pairwise sequence alignments can be effectively extended to derive multiple sequence alignments.)  This talk shall highlight the special role of pairwise kernel in feature selection and  SVM classification, embedded in a supervised wrapper approach.  This leads to an effective feature selection strategy called VIA-SVM.  In addition, as an application example, this talk will demonstrate how the proposed methods may be applied to (protein-sequence-based) sub-cellular localization analysis.

Bio of S.Y. Kung: