Data miners are usually confronted with a daunting array of variables from which they hope to discover useful relationships. One could always just test them individually, in groups, or in a stepwise procedure, using a sophisticated model similar or identical to that which the developer wants to ultimately deploy. This direct approach would usually be the best in the sense that it would discover the relationships that will ultimately be most useful.
Unfortunately, in most situations, this direct approach is much too costly in terms of computational resources. Training ...