INCREMENTALLY BUILDING ASPECT MODELS

• US 20080005137A1
• Filed: 06/29/2006
• Published: 01/03/2008
• Est. Priority Date: 06/29/2006
• Status: Abandoned Application

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Surendran, Arungunram C., Sra, Suvrit
• Application Number:
  US11/427,725
• Publication Number:
  US 20080005137A1
• Time in Patent Office: Days
• US Class Current: 707/101
• CPC Class Codes: G06F 18/2321 using statistics or function optimisation, e.g. modelling of probability density functions
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

The claimed subject matter relates to an unsupervised incremental learning framework, and in particular, to the creation and utilization of an unsupervised incremental learning framework that facilitates object discovery, clustering, characterization and/or grouping. Such an unsupervised incremental learning framework, once created, can thereafter be employed to incrementally estimate a latent variable model through the utilization of spectral and/or probabilistic models in order to incrementally cluster, discover, group and/or characterize tightly knit themes/topics within document sets and/or streams, thus leading to the generation of a set of themes/topics that better correlate with human perceptual labeling schemes.

Claims

1. A system that builds an aspect model from observed data, comprising:

an interface component that receives and incrementally extracts one or more aspects from the observed data; and

an analysis component that merges the one or more extracted aspects with an existing aspect model.

2. The system of claim 1, the analysis component determines whether the existing aspect model adequately describes the observed data.

3. The system of claim 2, when the analysis component ascertains that the existing aspect model inadequately describes the observed data, the analysis component indicates that additional aspects are needed.

4. The system of claim 1, the analysis component assigns weights to one or more parts of the observed data based at least in part on an adequacy determination on whether the existing aspect model describes the observed data.

6. The system of claim 1, the analysis component further comprising a scoring component that assigns a relevancy score to the observed data for each of the one or more extracted aspects.

7. The system of claim 6, the scoring component assigns probabilities to the observed data for each of the one or more extracted aspects.

8. The system of claim 6, the analysis component further comprising a weighting component that determines, based at least in part on the relevancy score and a mixing ratio, whether the one or more extracted aspects should be merged, deleted and/or assigned a reduced weight or an enhanced weight.

9. The system of claim 8, based at least in part on the relevancy score and the mixing ratio the analysis component eliminates existing aspects from the existing aspect model.

10. The system of claim 1, the one or more extracted aspects selected based at least in part by spectral methods.

11. The system of claim 1, the one or more extracted aspect selected based at least in part by a combination of probabilistic and spectral methods.

12. The system of claim 1, the interface component receives new data and the analysis component determines whether the new data is adequately described by the existing aspect model.

13. The system of claim 1, the interface component determines one or more stopping creation.

5. The system of claim 5, the analysis component assigns a low weight to observed data adequately described by the existing aspect model, and assigns a high weight to observed data inadequately described by the existing aspect model.

14. A method for building an aspect model from observed data, comprising:

employing a component to receive a stream of observed data;

incrementally extracting one or more aspects from the stream of observed data; and

adding the one or more extracted aspects to an existing aspect model.

15. The method of claim 14, the component employs a mechanism to transform the stream of observed data to a digital form.

16. The method of claim 15, the component selectively removes at least one stop word from the digital form.

17. The method of claim 14, the aspect model constructed utilizing a cost optimization function.

18. The method of claim 14, further comprising utilizing at least one statistical technique to evaluate a relative importance of the observed data in relation to every aspect included in the existing aspect model.

19. The method of claim 14, the adding further comprising utilizing an ascertained mixing ratio and a determined relevancy score.

20. A system that effectuates aspect model construction and notification, comprising:

means for continuously applying a dynamically expandable clustering series to a stream of data that comprises extractable aspects until an ascertainable stop condition is satisfied and extracting an extractable aspect;

means for receiving one or more user preferences;

means for ascertaining a match between the extractable aspect and the one or more user preferences; and

means for notifying one or more users of the match.