Inductive biases for stable feature selection in high dimensional spaces : applications to gene profiling and diagnosis from DNA microarrays
(2010)
Files
ThesisThibaultHelleputteHyperRef.pdf
Open Access - Adobe PDF
- 15.87 MB
Details
- Authors
- Supervisors
- Abstract
- (en) In many technological or industrial fields, the amount of high dimensional data is steadily growing. The number of dimensions is however often growing much faster than the number of points available. This setting makes many machine learning applications subject to the curse of dimensionality, making difficult the estimation of models that will generalize well. This thesis focuses on the design and application of techniques achieving highly sparse feature selection while allowing the estimation of models with good classification performance in a context where only few points are available, and those points lay in a high dimensional space. It deploys several means of regularization to mitigate the lack of extra samples in that high dimensional setting. It shows that this challenge can be successfully addressed provided that adequate inductive biases are used. One proposed approach is an ensemble method using internal information from the data only but taking many different « views » of the data. Other proposed approaches use external extra information. This extra information can be either expert prior knowledge, or contained in other datasets about related tasks (transfer learning or multi-task learning). All the proposed methods are tested over several gene expression microarray datasets for diagnosis and biomarker discovery tasks. Microarrays measure at once the rate of transcription (the expression) of thousands of genes into mRNA, the intermediate messengers between genes and proteins production. Those datasets are typically made of few tens of samples (patients) and thousands of dimensions (genes). A part of the thesis is dedicated to a case study, where an analysis of gene expression microarray data is performed in the context of the Cristall project. This project is an attempt to identify allergy risk factors in newborns.
- Affiliations
UCLouvain SST/ICTM - Institute for Information and Communication. Technologies, Electronics and Applied Mathematics
Citations
Helleputte, T. (2010). Inductive biases for stable feature selection in high dimensional spaces : applications to gene profiling and diagnosis from DNA microarrays. https://hdl.handle.net/2078.5/130955