Feature selection has become increasingly important in data mining in recent years due to the rapid increase in the dimensionality of big data. However, the reliability and consistency of feature selection methods (filters) vary considerably on different data and no single filter performs consistently well under various conditions. Therefore, feature selection ensemble has been investigated recently to provide more reliable and effective results than any individual one but all the existing feature selection ensemble treat the feature selection methods equally regardless of their performance. In this paper, we present a novel framework which applies weighted feature selection ensemble through proposing a systemic way of adding different weights to the feature selection methods-filters. Also, we investigate how to determine the appropriate weight for each filter in an ensemble. Experiments based on ten benchmark datasets show that theoretically and intuitively adding more weight to ‘good filters’ should lead to better results but in reality it is very uncertain. This assumption was found to be correct for some examples in our experiment. However, for other situations, filters which had been assumed to perform well showed bad performance leading to even worse results. Therefore adding weight to filters might not achieve much in accuracy terms, in addition to increasing complexity, time consumption and clearly decreasing the stability.
|Number of pages||7|
|Publication status||Published - 10 Nov 2015|
|Event||SAI Intelligent Systems Conference 2015 - London, United Kingdom|
Duration: 10 Nov 2015 → 11 Nov 2015
|Conference||SAI Intelligent Systems Conference 2015|
|Period||10/11/15 → 11/11/15|
- Feature selection