Methods and systems for analyzing discrete-valued datasets

• US 10,394,898 B1
• Filed: 09/15/2015
• Issued: 08/27/2019
• Est. Priority Date: 09/15/2014
• Status: Active Grant

• Current Assignee: The MathWorks, Inc.
• Original Assignee: The MathWorks, Inc.
• Inventors: Matteson, Andrew
• Primary Examiner(s): Vu, Bai D
• Application Number:
  US14/855,247
• Time in Patent Office: Days
• Field of Search: None
• CPC Class Codes: G06F 16/24578 using ranking
• CPC Class Codes: G06F 16/3347 using vector based model
• CPC Class Codes: G06F 16/93 Document management systems

Abstract

Methods and systems disclosed herein may be used to determine the structure of a dataset comprising discrete-valued data corresponding to features and items. In some embodiments, a device may receive a discrete-valued matrix with a first dimension corresponding to items and a second dimension corresponding to features. The device may calculate an engineered features set and a weights set for the matrix. The device may update the engineered features set using the weights set, and update the weights set using the updated engineered features set based on the mutual information between the matrix and one of the updated engineered features set. The device may receive a request indicating at least one of the engineered features set, identify items based on the matrix and the indicated at least one of the engineered features set, and provide a response based on the identified items.

Claims

1. A system for determining the structure of an electronic dataset, the system comprising:

one or more processors configured to perform the steps of;

receiving a matrix with a first dimension corresponding to items, a second dimension corresponding to features, and discrete-valued elements indicating a presence, absence, or frequency of the features in the items;

generating an engineered features set and a weights set for the matrix, the engineered features set and the weights set corresponding to latent structures in the matrix, generating the engineered features set and the weights set comprising;

generating a first engineered feature and a first weights vector corresponding to a first latent structure in the matrix, generating the first engineered feature and the first weights vector comprising;

updating the first engineered feature of the engineered features set using the matrix and the first weights vector of the weights set, and

updating the first weights vector of the weights set using a mutual information of the matrix and the first engineered feature; and

generating a second engineered feature and a second weights vector corresponding to a second latent structure using a subset of the matrix associated with the first latent structure, generating the second engineered feature and the second weights vector comprising;

determining the subset of the matrix using at least one of the first engineered feature and the first weights vector;

updating the second engineered feature of the engineered features set using the subset and the second weights vector of the weights set; and

updating the second weights vector of the weights set using a mutual information of the subset and the second engineered feature;

receiving a request indicating at least one of the engineered features set;

identifying items based on the matrix and the indicated at least one of the engineered features set; and

providing a response based on the identified items.

2. The system of claim 1, wherein the first engineered feature is updated and then the first weights vector is updated.

3. The system of claim 1, wherein the response comprises one or more of the identified items.

4. The system of claim 1, wherein the subset of the matrix depends on values of the first weights vector and a threshold value.

5. The system of claim 1, wherein the subset of the matrix depends on values of the first engineered feature and a threshold value.

6. The system of claim 1, wherein generating the engineered features set and the weights set further comprises generating clusters by clustering the items using at least the first engineered feature, and wherein the subset comprises rows of the matrix corresponding to items of a first one of the clusters.

7. The system of claim 6, the one or more processors further for associating an additional item with a second one of the clusters based on generated values of at least the first engineered feature for the additional item.

8. The system of claim 1, wherein:

a correlation between the first engineered feature and the matrix is negative, and

updating the first weights vector based on the mutual information of the matrix and the first engineered feature comprises calculating a negative of the mutual information of the first engineered feature and the matrix.

9. The system of claim 1, wherein the engineered features set includes a third engineered feature, and wherein generating the engineered features set and the weights set further comprises generating the third engineered feature and a third weights vector based on pre-categorized training data.

10. The system of claim 1, wherein generating the engineered features set and the weights set further comprises generating additional engineered features until a machine learning task performed using the first engineered feature and the additional engineered features satisfies a performance criterion.

11. A non-transitory computer-readable medium containing instructions that, when executed by one or more processors, cause a device to perform operations for determining the structure of an electronic dataset, the operations comprising:

receiving a discrete-valued dataset comprising rows corresponding to items and columns corresponding to features;

determining, based on the discrete-valued dataset;

a first engineered feature indicating a first latent structure in the discrete-valued dataset with elements of the first engineered feature corresponding to the items,

a first weights vector providing context for interpreting the first engineered feature with elements of the first weights vector corresponding to the features, and

wherein determining the first engineered feature comprises iteratively updating, until a condition is satisfied;

the elements of the first engineered feature based on a product of the first weights vector and rows of the discrete-valued dataset, and

the elements of the first weights vector based on a product of the first engineered feature and columns of the discrete-valued dataset;

generating a subset of the discrete-valued dataset based on at least the first engineered feature;

determining, using the subset of the discrete-valued dataset, a second engineered feature indicative of a second latent structure in the discrete-valued dataset and a second weights vector providing context for interpreting the second engineered feature; and

outputting indications of the first latent structure and the second latent structure.

12. The non-transitory computer-readable medium of claim 11, wherein the elements of the first engineered feature are updated and then the elements of the first weights vector are updated.

13. The non-transitory computer-readable medium of claim 11, the operations further comprising:

receiving a query item;

determining discrete values of the features for the query item;

determining values of the first engineered feature and the second engineered feature for the query item based on the first weights vector and the second weights vector;

determining one or more similar items in the discrete-valued dataset based on at least the first engineered feature and the second engineered feature; and

providing the one or more similar items to a user or non-transitory memory.

14. The non-transitory computer-readable medium of claim 11, wherein the subset of the discrete-valued dataset depends on the elements of the first weights vector and a threshold.

15. The non-transitory computer-readable medium of claim 11, wherein the subset of the discrete-valued dataset depends on the elements of the first engineered feature and a threshold.

16. The non-transitory computer-readable medium of claim 11, wherein generating the subset of the discrete-valued dataset based on at least the first engineered feature comprises clustering the items using at least the first engineered feature, and wherein the subset comprises rows of the discrete-valued dataset corresponding to items of a cluster.

17. The non-transitory computer-readable medium of claim 11, wherein:

a first element of the first weights vector corresponds to a first feature, and a value of the first element depends on a negative of a mutual information of the first engineered feature and a column of the discrete-valued dataset corresponding to the first feature; and

a correlation between the first engineered feature and the column of the discrete-valued dataset is negative.

18. The non-transitory computer-readable medium of claim 11, the operations further comprising calculating additional engineered features until a machine learning task performed using the first engineered feature and the additional engineered features satisfies a performance criterion.

19. A computer-implemented method for determining the structure of an electronic dataset, the method comprising:

receiving, on a computer, a first word corresponding to a term in a first term-document matrix for a first corpus of documents in a first language;

identifying a first engineered feature of the first term-document matrix corresponding to the first word;

determining a second engineered feature of a second term-document matrix corresponding to the first engineered feature, the second term-document matrix for a second corpus of documents in a second language;

identifying one or more second words in the second language corresponding to the second engineered feature; and

providing the identified one or more second words to a user or non-transitory memory.

20. The computer-implemented method of claim 19, wherein the first engineered feature and the second engineered feature correspond to a common latent structure.

21. The computer-implemented method of claim 19, wherein the first word concerns a concept in the first language and the identified one or more second words concern the concept in the second language.