Feature selection with a general hybrid algorithm

摘要: The Feature Selection problem involves discovering a subset of features, such that classifier built only with this would have better predictive accuracy than from the entire set features. A large number algorithms already been proposed for feature selection problem. Although significantly different regards to (1) the search strategy they use determine right features and (2) how each is evaluated, are usually classified in three general groups: Filters, Wrappers Hybrid solutions. In thesis, we propose new hybrid system machine learning. idea behind algorithm, FortalFS, extract combine best characteristics filters wrappers one algorithm. FortalFS uses results another as starting point through subsets evaluated by learning With an efficient heuristic, can decrease be consequently decreasing computational effort still able select accurate subset. We also designed variant original algorithm attempt work weighting order evaluate experiments were run compared well-known filter wrapper algorithms, Focus, Relief, LVF, others. Such aver datasets UCI Repository. Results showed outperforms most significantly. However, it presents time-consuming performance similar wrappers. Additional using specially artificial demonstrated identify remove both irrelevant, redundant randomly class-correlated The time-consumption issue addressed parallelism. parallel version based on master/slave design pattern implemented evaluated. In several experiments, achieve near optimal speedups.