Applied Mathematics Seminar - Machine Learning and Artificial Intelligence Event

Inference from Outliers

Classic machine learning is largely about classification and regression problems. However, many practical problems of interest in genomics, condition monitoring, medical diagnostics and security are better posed as problems of detecting novelty. In this talk, I will describe two applications of extracting useful information from novel data, in problems relating to modelling cellular protein concentrations and the solubility of synthetic chemical molecules. The algorithmic framework poses a robust support vector regression problem and the resulting non-convex optimisation problem is solved using a difference-of-convex formulation. I will further consider structured matrix approximation methods in the formulation of outlier detection and show how a robust version of principal component analysis, with sparsity constraints and a graph regulariser can be used in extracting outlier data, and illustrate the approach using functional genomics datasets.

(Joint work with: Yawwani Gunawardana, Greg Parkes, Pratheeba Jeyananthan and Omar Shetta)