I N F L U E N C E V A L U E S R E P O R T order attributes by influence values = true ============================================= AutoClass CLASSIFICATION for the 100 cases in /home/wtaylor/AC/autoclass-c/data/3-dim/3-dimc.db2 /home/wtaylor/AC/autoclass-c/data/3-dim/3-dimc.hd2 with log-A (approximate marginal likelihood) = -4255.623 from classification results file /home/wtaylor/AC/autoclass-c/data/3-dim/3-dimc.results-bin and using models /home/wtaylor/AC/autoclass-c/data/3-dim/3-dimc.model - index = 0 ORDER OF PRESENTATION: * Summary of the generating search. * Weight ordered list of the classes found & class strength heuristic. * List of class cross entropies with respect to the global class. * Ordered list of attribute influence values summed over all classes. * Class listings, ordered by class weight. _____________________________________________________________________________ _____________________________________________________________________________ SEARCH SUMMARY 4 tries over 2 seconds _______________ SUMMARY OF 10 BEST RESULTS _________________________ ## PROBABILITY exp(-4255.623) N_CLASSES 2 FOUND ON TRY 4 *SAVED* -1 PROBABILITY exp(-4257.323) N_CLASSES 2 FOUND ON TRY 2 *SAVED* -2 PROBABILITY exp(-4257.496) N_CLASSES 2 FOUND ON TRY 1 PROBABILITY exp(-4257.508) N_CLASSES 2 FOUND ON TRY 3 ## - report filenames suffix _____________________________________________________________________________ _____________________________________________________________________________ CLASSIFICATION HAS 2 POPULATED CLASSES (max global influence value = 0.172) We give below a heuristic measure of class strength: the approximate geometric mean probability for instances belonging to each class, computed from the class parameters and statistics. This approximates the contribution made, by any one instance "belonging" to the class, to the log probability of the data set w.r.t. the classification. It thus provides a heuristic measure of how strongly each class predicts "its" instances. Class Log of class Relative Class Normalized num strength class strength weight class weight 0 -4.21e+01 1.00e+00 58 0.584 1 -4.30e+01 4.31e-01 42 0.416 CLASS DIVERGENCES The class divergence, or cross entropy w.r.t. the single class classification, is a measure of how strongly the class probability distribution function differs from that of the database as a whole. It is zero for identical distributions, going infinite when two discrete distributions place probability 1 on differing values of the same attribute. Class class cross entropy Class Normalized num w.r.t. global class weight class weight 0 2.86e-01 58 0.584 1 5.16e-01 42 0.416 ORDERED LIST OF NORMALIZED ATTRIBUTE INFLUENCE VALUES SUMMED OVER ALL CLASSES This gives a rough heuristic measure of relative influence of each attribute in differentiating the classes from the overall data set. Note that "influence values" are only computable with respect to the model terms. When multiple attributes are modeled by a single dependent term (e.g. multi_normal_cn), the term influence value is distributed equally over the modeled attributes. num description I-*k 000 X_coordinate 1.000 001 Y_coordinate 1.000 002 Z_coordinate 1.000 CLASS LISTINGS: These listings are ordered by class weight -- * j is the zero-based class index, * k is the zero-based attribute index, and * l is the zero-based discrete attribute instance index. Within each class, the covariant and independent model terms are ordered by their term influence value I-jk. Covariant attributes and discrete attribute instance values are both ordered by their significance value. Significance values are computed with respect to a single class classification, using the divergence from it, abs( log( Prob-jkl / Prob-*kl)), for discrete attributes and the relative separation from it, abs( Mean-jk - Mean-*k) / StDev-jk, for numerical valued attributes. For the SNcm model, the value line is followed by the probabilities that the value is known, for that class and for the single class classification. Entries are attribute type dependent, and the corresponding headers are reproduced with each class. In these -- * num/t denotes model term number, * num/a denotes attribute number, * t denotes attribute type, * mtt denotes model term type, and * I-jk denotes the term influence value for attribute k in class j. This is the cross entropy or Kullback-Leibler distance between the class and full database probability distributions (see interpretation-c.text). * Mean StDev -jk -jk The estimated mean and standard deviation for attribute k in class j. * |Mean-jk - The absolute difference between the Mean-*k|/ two means, scaled w.r.t. the class StDev-jk standard deviation, to get a measure of the distance between the attribute means in the class and full data. * Mean StDev The estimated mean and standard -*k -*k deviation for attribute k when the model is applied to the data set as a whole. * Prob-jk is known 1.00e+00 Prob-*k is known 9.98e-01 The SNcm model allows for the possibility that data values are unknown, and models this with a discrete known/unknown probability. The gaussian normal for known values is then conditional on the known probability. In this instance, we have a class where all values are known, as opposed to a database where only 99.8% of values are known. CLASS 0 - weight 58 normalized weight 0.584 relative strength 1.00e+00 ******* class cross entropy w.r.t. global class 2.86e-01 ******* Model file: /home/wtaylor/AC/autoclass-c/data/3-dim/3-dimc.model - index = 0 Numbers: numb/t = model term number; numb/a = attribute number Model term types (mtt): (multi_normal_cn MNcn) REAL ATTRIBUTE (t = R) |Mean-jk - numb t mtt description I-jk Mean StDev Mean-*k|/ Mean StDev t a -jk -jk StDev-jk -*k -*k 00 01 R MNcn Y_coordinate ....... 0.095 ( 5.46e+00 4.51e+00) 6.00e-01 ( 2.75e+00 5.66e+00) 00 00 R MNcn X_coordinate ....... 0.095 ( 5.31e+00 8.21e+00) 2.00e-01 ( 3.67e+00 9.09e+00) 00 02 R MNcn Z_coordinate ....... 0.095 (-8.48e-02 8.05e-01) 1.82e-01 ( 6.14e-02 9.77e-01) Correlation matrix (row & column indices are attribute numbers) 0 1 2 0 1.000 0.777 0.298 1 0.777 1.000 0.203 2 0.298 0.203 1.000 CLASS 1 - weight 42 normalized weight 0.416 relative strength 4.31e-01 ******* class cross entropy w.r.t. global class 5.16e-01 ******* Model file: /home/wtaylor/AC/autoclass-c/data/3-dim/3-dimc.model - index = 0 Numbers: numb/t = model term number; numb/a = attribute number Model term types (mtt): (multi_normal_cn MNcn) REAL ATTRIBUTE (t = R) |Mean-jk - numb t mtt description I-jk Mean StDev Mean-*k|/ Mean StDev t a -jk -jk StDev-jk -*k -*k 00 01 R MNcn Y_coordinate ....... 0.172 (-1.07e+00 4.95e+00) 7.71e-01 ( 2.75e+00 5.66e+00) 00 00 R MNcn X_coordinate ....... 0.172 ( 1.35e+00 1.00e+01) 2.31e-01 ( 3.67e+00 9.09e+00) 00 02 R MNcn Z_coordinate ....... 0.172 ( 2.68e-01 1.18e+00) 1.75e-01 ( 6.14e-02 9.77e-01) Correlation matrix (row & column indices are attribute numbers) 0 1 2 0 1.000 0.838 0.067 1 0.838 1.000 0.171 2 0.067 0.171 1.000