High-dimensional Feature Selection in Classification: A Length-Adaptive Evolutionary Approach

Feature selection is an essential technique which has been widely applied in data mining. Recent research has shown that a good feature subset can be obtained by using evolutionary computing (EC) approaches as a wrapper. However, most feature selection methods based on EC use a fixed-length encoding to represent feature subsets. When this fixed length representation is applied to high-dimensional data, it requires a large amount of memory space as well as a high computational cost. Moreover, this representation is inflexible and may limit the performance of EC because of a too huge search space. In this paper, we propose an Adaptive- Variable-Length Genetic Algorithm (A VLGA), which adopts a variable-length individual encoding and enables individuals with different lengths in a population to evolve in their own search space. An adaptive length changing mechanism is introduced which can extend or shorten an individual to guide it to explore in a better search space. Thus, A VLGA is able to adaptively concentrate on a smaller but more fruitful search space and yield better solutions more quickly. Experimental results on 6 high-dimensional datasets reveal that A VLGA performs significantly better than existing methods.