Feature Sensitivity Analysis

Ideally, when determining the optimisal classification, the best features are added first and the less useful ones added later. This results in a decreasing rate at which the probability of misclassification (PMC) decreases. Unfortunately, the addition of new features involves measuring new parameters. Due to inexact parameter estimation, a point is reached at which the addition of further features actually decreases classification performance. This occurs when the increase in error caused by measurement error is greater than the decrease in PMC achieved by adding the new feature. The number of features which produces the lowest expected PMC is called the optimal number of features.

Because features are not usually ranked according to their effectiveness, sample information must be used to compare and rank them. These estimates are not exact and so the features cannot be ranked properly. The problem of determining the optimal subset of features for empirically ordered features is unsolved except for an exhaustive search. Suboptimal methods do exist which require less computation. Those provided here are derived from NumSelectC.

Subtopics:

Classes: