B0639
Title: Supervised projection pursuit for discriminant analysis through machine learning
Authors: Donald Jacobs - University of North Carolina at Charlotte (United States) [presenting]
Tyler Grear - University of North Carolina at Charlotte (United States)
Chris Avery - University of North Carolina at Charlotte (United States)
Abstract: Supervised projection pursuit is implemented as a neural network used for dimensionality reduction and building classification models. The architecture consists of a single layer of interacting neurons that serve as both input and output layers. Each neuron has access to the mean and covariance of the input data, and represents a basis vector that projects data to obtain two emergent features (mean and variance). Efficacy measures how well data clusters in a mode feature space plane. Because efficacy is linearly separable across all projections, a stochastic process prioritizes Jacobi or Cayley rotations to monotonically increase net efficacy, while the complete and orthonormal basis set is maintained. Each $k$-th projection contributes to the efficacy associated with a rectifying unit as the product of two conjugate variables: $E(k) = Q(k)S(k)$, where $E(k)$ is efficacy, $S(k)$ is a projection index, and $Q(k)$ is the quality of clustering. Once converged, multivariate differences and similarities are quantified by projections within discriminant and indifferent subspaces. The application to molecular function recognition is described. The input is a time series of atomic motions over many molecular trajectories of functional and non-functional molecules. The results identify functional dynamics that describe biochemical mechanisms important for drug discovery.
