1 Reply Latest reply: Aug 16, 2005 4:30 PM by 396396 RSS

    Attribute Importance


      in which range the importance value has to be,
      to call the predictor a good predictor?

      The highest value of the result I've got is 0,09.

      Thanks a lot

        • 1. Re: Attribute Importance
          The importance value is an MDL measure, the mean per row compression in bits. Thus the significance of the compression increases with the number of rows. For a binary problem with balanced data, the number of bits per row is 1, so that .09 is a 9% reduction in communication cost. With highly skewed data, say 95%-5%, the average bits per row is much less ~.28 bits, thus the reduction is approximately 35%.