Dominance Pruning in Machine Learning for Solving Financial Trading and Real-Time Multimedia Applications

Abstract

This paper presents the design of dominance relations to reduce the space traversed in machine learning for solving two applications in financial trading and real-time multimedia. A machine-learning algorithm designed for an application with a huge search space will need to perform an efficient traversal of the space during learning. It will be more effective if it employs a powerful pruning mechanism to eliminate suboptimal candidates before using them in the learning algorithm. In our approach, we present dominance relations for pruning subspaces with suboptimal kernels that are otherwise evaluated in learning, where kernels represent the statistical quality, average density, or probability of solutions in a subspace. Specifically, when one subspace dominates another by a dominance relation, we can prune the latter and guarantee without searching both that the kernel of the latter cannot be better than that of the first. As a result, a significant portion of the search space will be pruned by those non- dominated subspaces during learning. In the financial trading application studied, we use mean reversion as our strategy for learning the set of promising stocks and Pareto-optimality as our dominance relation to reduce the space evaluated in learning. In the multimedia application, we propose a dominance relation using an axiom from our past work to approximate the subspace of perceptual qualities within an error threshold. The pruning mechanism allows the learning of the mapping from controls to perceptual qualities while eliminating the evaluation of all those mappings that are within the error thresholds. In both cases, we can harness the complexity of machine learning by reducing the candidate space evaluated.