Method of and apparatus for automated behavior prediction

• US 8,010,470 B2
• Filed: 11/24/2009
• Issued: 08/30/2011
• Est. Priority Date: 04/15/2005
• Status: Expired due to Fees

• Current Assignee: Leidos, Inc. (Leidos Holdings, Inc.)
• Original Assignee: Science Applications International Corporation
• Inventors: Jackson, Gary M.
• Primary Examiner(s): Gaffin; Jeffrey A
• Assistant Examiner(s): Wong; Lut
• Application Number:
  US12/625,087
• Publication Number:
  US 20100161528A1
• Time in Patent Office: 644 Days
• Field of Search: 706/20, 706/21, 706/45
• US Class Current: 706/20
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising

Abstract

A computer-implemented method of behavior prediction includes selecting behavior examples having corresponding antecedent candidates, identifying source text descriptions describing the behavior examples, automatically extracting predictors as common themes across all statements and all behavior examples with a language-independent theme extraction process, flagging each behavior example to indicate a presence or absence of the corresponding extracted antecedents in each of the source text descriptions and creating a data array consisting of antecedent columns and behavior example rows, submitting the data array to a pattern classifier to extract patterns among the antecedent candidates and outcomes by training and validating the pattern classifier and predicting a new occurrence of a target behavior by entering a current state of the antecedents to the trained pattern classifier.

Claims

1. A non-transitory computer-readable medium comprising computer-executable instructions for behavior prediction, the computer-executable instructions comprising:

a formatter that splits source text descriptions describing behavior examples, and associated antecedent candidates into text segments and tags the text segments with unique identifiers identifying the source text descriptions from which the text segments were obtained;

an automated antecedent theme extractor that;

creates antecedents by cross-correlating words contained in the text segments to create a list of word pairs and cross-correlating each of the word pairs to one or more other word pairs in the list to create pair-pairs, such that the antecedents comprise a word pair from the list of word pairs, pair-pairs corresponding to the word pair, and any text segment containing the word pair;

eliminates antecedents that have fewer than a minimum number of text segments listed thereunder, and

generates an antecedent and outcome data array containing rows corresponding to the text segments, columns corresponding to the antecedents and containing data indicating whether or not the antecedents are present in the text segments, and columns corresponding to outcomes and containing data indicating whether or not the outcomes are present in the text segments; and

a pattern classifier that receives the antecedent and outcome data array as a training input and outputs a training output in response to the data for each text segment included in the antecedent and outcome data array, and receives as a prediction input a current antecedent presence-absence vector indicating a presence or absence of the antecedents included in the antecedent and outcome data array in a current environment and outputs a prediction of a new occurrence of a target behavior in response to the current antecedent presence-absence vector.

2. The non-transitory computer-readable medium of claim 1, wherein the computer-executable instructions further comprise a validation unit that analyzes an accuracy of the training output from the pattern classifier and trains the pattern classifier to improve an accuracy of the training output until a minimum training error is achieved.

3. The non-transitory computer-readable medium of claim 2, wherein the validation unit trains the pattern classifier before the pattern classifier receives the current antecedent presence-absence vector as a prediction input.

4. The non-transitory computer-readable medium of claim 2, wherein the validation unit trains the pattern classifier using a leaving-one-out cross-validation methodology.

5. The non-transitory computer-readable medium of claim 1, wherein the pattern classifier is a BPN (back-propagation neural network) pattern classifier.

6. The non-transitory computer-readable medium of claim 1, wherein the automated antecedent theme extractor processes the text segments in a language-independent manner.

7. A computer-implemented method of behavior prediction, comprising:

downloading from a database of behavior examples a list of antecedent candidates associated with the behavior examples with unique identifiers identifying the behavior examples;

processing the list of antecedent candidates with the unique identifiers with an automated antecedent theme extractor that;

creates antecedents by cross-correlating words contained in the antecedent candidates to create a list of word pairs and cross-correlating each of the word pairs to one or more other word pairs in the list to create pair-pairs, such that the antecedents comprise a word pair from the list of word pairs, pair-pairs corresponding to the word pair, and any antecedent candidate containing the word pair;

eliminates antecedents that have fewer than a minimum number of antecedent candidates listed thereunder, and

generating an antecedent and outcome data array containing rows corresponding to the behavior examples, columns corresponding to the antecedents and containing data indicating whether or not the antecedents are present in the behavior examples, and columns corresponding to outcomes and containing data indicating whether or not the outcomes are present in the text segments;

processing the antecedent and outcome data array as a training input with a pattern classifier that outputs a training output in response to the data for each behavior example included in the antecedent and outcome data array;

analyzing an accuracy of the training output from the pattern classifier with a validation unit that trains the pattern classifier to improve an accuracy of the training output until a minimum training error is achieved; and

processing a current antecedent presence-absence vector indicating a presence or absence of the antecedents included in the antecedent and outcome data array in a current environment as a prediction input with the pattern classifier trained by the validation unit that outputs a prediction of a new occurrence of a target behavior in response to the current antecedent presence-absence vector.

8. The computer-implemented method of claim 7, wherein the validation unit trains the pattern classifier using a leaving-one-out cross-validation methodology.

9. The behavior predictor of claim 7, wherein the pattern classifier is a BPN (back-propagation neural network) pattern classifier.

10. The behavior predictor of claim 7, wherein automated antecedent theme extractor processes the list of the antecedent candidates with the unique identifiers in a language-independent manner.