Variable Importance Scores

There are many methods of scoring the importance variables in prediction a response but not much is known about their accuracy. This paper partially fills gap by introducing new method based on GUIDE algorithm and comparing it with 11 existing methods. For data without missing values, eight shown to give biased scores that too high or low, depending type (ordinal, binary nominal) whether they dependent other variables, even when all them independent response. Among remaining four methods, only continues unbiased if there values. It does this self-calibrating bias-correction step applicable also provides threshold for differentiating important from unimportant 95 99 percent confidence. Correlations predictive power studied three real sets. correlations marginal higher than conditional power.