Data spiders
AssigneeFair Isaac Corporation
InventorsHelen Geraldine E. Rosario, Michael S. Sossi, Christopher Ralph
Life of this patent
Estimating the term…
Abstract
Data spiders, provide an automated system that can take a file or file store of historic transaction data and create the best set of variables from that data, where “best” means highly predictive. Genetic algorithms are used to parameterized transactions to form groups, which are subjected naïve Bayes score ranking. Variable groups are generated and ranked accord to the score. Data spiders span the full information available, are uncorrelated with previous methods, and are easily interpretable.
Claims23
A computer implemented method of data mining transactional data through data spiders implementing genetic algorithms, comprising the steps of:
- programmatically creating groups of variables from parameterized transactional data;
- using a naïve Bayes model to calculate score cards for each group;
- determining group divergence by naïve Bayes scores and compiling scorecards quantifying the divergence, wherein divergence measures the ability of a naïve Bayes score to separate two outcome classes of a binary target variable;
- determining fitness of each group by the quantified divergence of the naïve Bayes score;
- ranking the groups by fitness, from divergence score;
- identifying elite groups based on top ranking groups;
- iteratively performing the following sequence of steps until the fitness measure of the pre-set top ranking number of groups does not substantively change, those steps comprising:
- selecting parent groups based on fitness, wherein parent groups are cloned elites from a previous iteration;
- creating children groups, wherein each children is a present iteration variable group combination of two parent groups;
- mutating children groups, wherein each mutating group comprises a pre-set fraction of children groups with at least one altered parameter value per group;
- merging the parent groups and children groups;
- assigning a fitness value of each group based on divergence value;
- ranking the groups by fitness values;
- and
- tagging the top ranking groups as elite groups to be promoted to the next iteration;
- wherein groups of transaction variables using analogous genetic processes, iteratively valued, scored, and ranked for a particular mathematical score of fitness, progressively maintaining top ranking groups while introducing mutating groups results in an automated process of comprehensively mining the transaction data for predictors of a specified target objective;
- and outputting said predictors,
A computer program residing on a computer-readable media, said computer program for directing an informational technology system to systematically data mine transactional data through data spiders implementing genetic algorithms, comprising the steps of:
- programmatically creating groups of variables from parameterized transactional data;
- using a naïve Bayes model to calculate score cards for each group;
- determining group divergence by naïve Bayes scores and compiling scorecards quantifying the divergence, wherein divergence measures the ability of a naïve Bayes score to separate two outcome classes of a binary target variable;
- determining a fitness value of at least two groups by the quantified divergence of the naïve Bayes score;
- ranking the groups by fitness values, from highest to lowest divergence;
- identifying elite groups based on top ranking groups;
- iteratively performing the following sequence of steps until the fitness measure of the pre-set top ranking number of groups does not change substantively, those steps comprising:
- selecting parent groups based on fitness, wherein parent groups are cloned elites from a previous iteration;
- determining and creating children groups, wherein at least two children groups comprise the present iteration variable group combination of two parent groups;
- determining and mutating children groups, wherein at least one mutating group is a pre-set fraction of children groups with at least one altered parameter value per group;
- merging the parent groups and children groups;
- determining and assigning fitness values of at least two groups;
- ranking the groups by fitness values;
- and
- tagging the top ranking groups as elite groups to be promoted to the next iteration without further changes;
- wherein groups of transaction variables using analogous genetic processes, iteratively valued, scored and ranked for a particular mathematical score of fitness, progressively maintaining top ranking groups while introducing mutating groups results in an automated process of comprehensively mining the transaction data for predictors of a specified target objective;
- and outputting said predictors.
Dependencies are inferred from the claim text.