ROMJIST Volume 22, No. 1, 2019, pp. 85-99
Camil Băncioiu, Maria Vințan, Lucian Vințan Efficiency Optimizations for Koller and Sahami's Feature Selection Algorithm
ABSTRACT: This article describes and evaluates four optimizations for Koller and Sahami's Feature Selection algorithm, significantly reducing the time it requires to complete. The optimizations exploit the Information Theory concepts used by the algorithm, its inherent data parallelism and the fact that much of the calculations it performs are redundant. Each proposed optimization was carefully evaluated, showing significant efficiency gains. In particular, a decomposition of conditional mutual information is shown to reduce the time required to calculate its primary heuristic and can be potentially applied to other algorithms which calculate conditional mutual information.KEYWORDS: Feature Selection, Koller and Sahami’s Algorithm, Markov Blankets, Conditional Mutual Information, Data Parallelism, Caching Techniques, Computation ReuseRead full text (pdf)