Feature selection method, device and apparatus for constructing machine learning model

Patent metadata

Field	Value
Publication number	US-11222285-B2
Application number	US-202117162939-A
Country	US
Kind code	B2
Filing date	Jan 29, 2021
Priority date	Oct 24, 2018
Publication date	Jan 11, 2022
Grant date	Jan 11, 2022

Abstract

A feature-selection system obtains a training data set and associated features; divides the training data set into a first number of training data subsets; and forms a plurality of feature-selecting data sets. A feature-selecting data set comprises a second number of training data subsets. The system processes, in parallel, each feature-selecting data set, which comprises: computing a first evaluation index for the features based on the feature-selecting data set; obtaining a group of index ranks corresponding to the features based on the first evaluation index; and obtaining a group of importance ranks corresponding to the features based on the feature-selecting data set and a machine-learning model. The system further obtains a group of total ranks by fusing groups of index ranks and groups importance ranks obtained from processing the plurality of feature-selecting data sets; and selecting target features from the features based on the group of total ranks.

First claim

What is claimed is:

1. A computer-executed method, comprising:
obtaining a training data set and features associated with the training data set;
dividing the training data set into a first number of training data subsets;
forming a plurality of feature-selecting data sets, wherein a respective feature-selecting data set comprises a second number of training data subsets, the second number being smaller than the first number;
processing, by a computer in parallel, each feature-selecting data set, which comprises:
computing a first evaluation index for the features based on the feature-selecting data set;
obtaining a group of index ranks corresponding to the features based on the first evaluation index; and
obtaining a group of importance ranks corresponding to the features based on the feature-selecting data set and a machine-learning model;

obtaining a group of total ranks corresponding to the features by fusing groups of index ranks and groups importance ranks obtained from processing the plurality of feature-selecting data sets; and
selecting target features from the features based on the group of total ranks.

2. The method according to claim 1 , wherein obtaining the group of total ranks of the features comprises:
fusing the groups of index ranks obtained from processing the plurality of feature-selecting data sets to obtain a group of total index ranks;
fusing the groups of importance ranks obtained from processing the plurality of feature-selecting data sets to obtain a group of total importance ranks; and
fusing the group of total index ranks and the group of total importance ranks to obtain the group of total ranks of the plurality of features.

3. The method according to claim 2 , wherein processing each feature-selecting data set further comprises:
computing additional evaluation indices for the features based on the feature-selecting data set; and
obtaining additional groups of index ranks corresponding to the features based on the additional evaluation indices.

4. The method according to claim 3 , wherein fusing the groups of index ranks to obtain the group of total index ranks of the plurality of features comprises:
for each evaluation index selected from a set of evaluation indices comprising the additional evaluation indices and the first evaluation index:
extracting from groups of index ranks association with the set of evaluation indices and the plurality of feature-selecting data sets, a plurality of groups of index ranks associate with the selected evolution index; and
performing a first rank fusion operation to respectively fuse corresponding ranks of the features in the extracted plurality of groups of index ranks to obtain a group of comprehensive index ranks corresponding to the selected evaluation index; and

performing a second rank fusion operation to respectively fuse groups of index ranks obtained for the set of evaluation indices to obtain the total index ranks of the features.

5. The method according to claim 2 , wherein the first rank fusion operation or the second rank fusion operation comprises one of: a mean operation, a maximum operation, a minimum operation, a weighted average operation and a robust rank aggregation (RRA) operation.

6. The method according to claim 1 , wherein dividing the training data set comprises one of:
dividing the training data set based on time; and
dividing the training data set randomly.

7. The method according to claim 1 , wherein the first evaluation index comprises one of: an information value (IV), a Gini coefficient (GINI), an information gain (IG), mutual information (MI), a Relief score, and a sample stability index (PSI).

8. The method according to claim 1 , wherein the training data set is divided into k training data subsets, wherein k different feature-selecting data sets are formed, and wherein each feature-selecting data set comprises k-1 training data subsets.

9. A computer system, comprising:
a processor;
a storage device coupled to the processor and storing instructions, which when executed by the processor cause the processor to perform a method, the method comprising:
obtaining a training data set and features associated with the training data set;
dividing the training data set into a first number of training data subsets;
forming a plurality of feature-selecting data sets, wherein a respective feature-selecting data set comprises a second number of training data subsets, the second number being smaller than the first number;
processing in parallel, each feature-selecting data set, which comprises:
computing a first evaluation index for the features based on the feature-selecting data set;
obtaining a group of index ranks corresponding to the features based on the first evaluation index; and
obtaining a group of importance ranks corresponding to the features based on the feature-selecting data set and a machine-learning model;

obtaining a group of total ranks corresponding to the features by fusing groups of index ranks and groups importance ranks obtained from processing the plurality of feature-selecting data sets; and
selecting target features from the features based on the group of total ranks.

10. The computer system according to claim 9 , wherein obtaining the group of total ranks of the features comprises:
fusing the groups of index ranks obtained from processing the plurality of feature-selecting data sets to obtain a group of total index ranks;
fusing the groups of importance ranks obtained from processing the plurality of feature-selecting data sets to obtain a group of total importance ranks; and
fusing the group of total index ranks and the group of total importance ranks to obtain the group of total ranks of the plurality of features.

11. The computer system according to claim 10 , wherein processing each feature-selecting data set further comprises:
computing additional evaluation indices for the features based on the feature-selecting data set; and
obtaining additional groups of index ranks corresponding to the features based on the additional evaluation indices.

12. The computer system according to claim 11 , wherein fusing the groups of index ranks to obtain the group of total index ranks of the plurality of features comprises:
for each evaluation index selected from a set of evaluation indices comprising the additional evaluation indices and the first evaluation index:
extracting from groups of index ranks association with the set of evaluation indices and the plurality of feature-selecting data sets, a plurality of groups of index ranks associate with the selected evolution index; and
performing a first rank fusion operation to respectively fuse corresponding ranks of the features in the extracted plurality of groups of index ranks to obtain a group of comprehensive index ranks corresponding to the selected evaluation index; and

performing a second rank fusion operation to respectively fuse groups of index ranks obtained for the set of evaluation indices to obtain the total index ranks of the features.

13. The computer system according to claim 10 , wherein the first rank fusion operation or the second rank fusion operation comprises one of: a mean operation, a maximum operation, a minimum operation, a weighted average operation and a robust rank aggregation (RRA) operation.

14. The computer system according to claim 9 , wherein dividing the training data set comprises one of:
dividing the training data set based on time; and
dividing the training data set randomly.

15. The computer system according to claim 9 , wherein the first evaluation index compris

Assignees

• Advanced New Technologies Co Ltd

Inventors

• Tang Yuzhou
• Jin Hong
• Wang Weiqiang
• Zhao Wenbiao

Classifications

• G06N20/20

  Ensemble learning · CPC title

• G06F16/24578

  using ranking · CPC title

• G06N99/00 Primary

  Subject matter not provided for in other groups of this subclass · CPC title

• G06F16/2228 Primary

  Indexing structures · CPC title

• G06N20/00 Primary

  Machine learning · CPC title