On the Complexity and Recognition Rates in Feature Selection.
SOUTHEASTERN MASSACHUSETTS UNIV NORTH DARTMOUTH
Pagination or Media Count:
For any pattern recognition problem, it is necessary to consider a meaningful constraint on the computational complexity. This is particularly true in selecting the optimum feature subset. Here complexity includes both the computation time and the memory space. Of course, the recognition rate, i.e. the probability of correct recognition, is another constraint which must simultaneously be taken into account. As the recognition rate reaches certain level, further improvement may require excessive amount of computational complexity. Thus for any specific application, there must be a trade-off between the two factors. Exhaustive search for the optimum feature subset is not practical. Sequential recursive procedure is shown to provide the best compromise in complexity and recognition rate for the construction of the best feature subset. Recognition rates for feature subsets selected by Bhattacharyya distance, mutual information and direct error probability estimation are also compared. With both the complexity and recognition rates in mind, simple and effective feature are extracted from the ACDA seismic data for realtime seismic discrimination. Modified author abstract