Noise mitigation in vector space representations of item collections

• US 10,339,407 B2
• Filed: 04/18/2017
• Issued: 07/02/2019
• Est. Priority Date: 04/18/2017
• Status: Active Grant

• Current Assignee: Maxim Analytics, LLC
• Original Assignee: Maxim Analytics, LLC
• Inventors: Bradford, Roger
• Primary Examiner(s): Kuddus, Daniel A
• Application Number:
  US15/490,722
• Publication Number:
  US 20180300580A1
• Time in Patent Office: 805 Days
• CPC Class Codes: G06F 18/2113 by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06V 10/30 Noise filtering
• CPC Class Codes: G06V 10/771 Feature selection, e.g. selecting representative features from a multi-dimensional feature space

Abstract

Mitigation of noise in vector space representations of item collections for applications such as facial recognition and spam detection. After receiving a collection of items, each item characterized by a plurality of features, a first vector space representative of the collection is prepared. Each item in the first vector space is represented by a vector that is a function of the features characterizing the item. Based on information characteristic of the first vector space, at least one of items contributing to first vector space noise, and features contributing to first vector space noise are identified. The first identified items and features are weighted to decrease the influence of the first identified items and features on vector spaces representative of the collection. A second vector space representative of the collection is prepared based on the first weighted items and features.

Claims

1. A computer-implemented method to mitigate noise in vector space representations of item collections, the method comprising:

receiving, by one or more computing devices, a collection of items, each item characterized by a plurality of features;

preparing, by the one or more computing devices, a first vector space representative of the collection, wherein each item in the first vector space is represented by a vector that is a function of the features characterizing the item;

first identifying, by the one or more computing devices and based on information characteristic of the first vector space, only a first one of;

items contributing to first vector space noise, and features contributing to first vector space noise;

first weighting, by the one or more computing devices, the first identified items and features to decrease the influence of the first identified items and features on vector spaces representative of the collection;

preparing, by the one or more computing devices, a second vector space representative of the collection based on the first weighted items and features;

second identifying, based on information characteristic of the second vector space, the first one of;

items contributing to second vector space noise and

features contributing to second vector space noise;

second weighting the second identified items and features to decrease the influence of the second identified items and features on vector spaces representative of the collection; and

preparing a third vector space representative of the collection based on the cumulatively weighted items and features, wherein each item is represented by a vector that is sum of the vectors of the features comprising the item.

2. The method of claim 1, wherein the first weighting comprises deleting the first identified items from the collection prior to preparing the second vector space.

3. The method of claim 1, wherein:

each item is characterized by a category feature having a category value; and

first identifying items comprises identifying particular items characterized by a vector in the first vector space that is closer to the centroid of vectors associated with at least one other category value than to the centroid of the vectors associated with the category value of the particular item.

4. The method of claim 1, wherein:

each item is characterized by a category feature having a category value, and each category value is represented by the centroid of items characterized by the category value; and

first identifying features contributing to first vector space noise comprises;

identifying the N feature vectors closest to the centroid of the item vectors of at least one category value; and

identifying the M feature vectors farther than X feature vectors from the centroid of the item vectors of at least one category value, wherein X is greater than or equal to N; and

first weighting the first identified items and features comprises;

weighting the identified N feature vectors to increase the influence of the N feature vectors on vector spaces representative of the collection, and

weighting the identified M feature vectors to decrease the influence of the M feature vectors on vector spaces representative of the collection.

5. The method of claim 1:

wherein first identifying comprises identifying only a first one of;

items contributing to vector space noise and

features contributing to vector space noise; and

the method further comprises, after preparing the second vector space representation;

second identifying, based on information characteristic of the second vector space, the other one of;

items contributing to second vector space noise and

features contributing to second vector space noise;

second weighting the second identified items and features to decrease the influence of the second identified items and features on vector spaces representative of the collection; and

preparing a third vector space representative of the collection based on the cumulatively weighted items and features, wherein each item is represented by a vector that is a function of the features comprising the item.

6. The method of claim 1, wherein each prepared vector space is a latent semantic indexed vector space.

7. The method of claim 1:

wherein each item is characterized by a category feature having a category value; and

the method further comprises;

folding in to the second vector space a vector corresponding to an item to be categorized; and

categorizing the item to be categorized with the category value of the category feature having a vector in the second vector space closest to the folded-in vector.

8. A computer program product, comprising:

a non-transitory computer-readable storage device having computer-executable program instructions embodied thereon that when executed by a computer cause the computer to mitigate noise in vector space representations of item collections, the computer-executable program instructions comprising;

computer-executable program instructions to receive a collection of items, each item characterized by a plurality of features;

computer-executable program instructions to prepare a first vector space representative of the collection, wherein each item in the first vector space is represented by a vector that is a function of the features characterizing the item;

computer-executable program instructions to first identify, based on information characteristic of the first vector space, only a first one of;

items contributing to first vector space noise, and

features contributing to first vector space noise;

computer-executable program instructions to first weight the first identified items and features to decrease the influence of the first identified items and features on vector spaces representative of the collection;

computer-executable program instructions to prepare a second vector space representative of the collection based on the first weighted items and features; and

computer-executable program instructions to second identify, based on information characteristic of the second vector space, the first one of;

items contributing to second vector space noise and

features contributing to second vector space noise;

computer-executable program instructions to second weight the second identified items and features to decrease the influence of the second identified items and features on vector spaces representative of the collection; and

computer-executable program instructions to prepare a third vector space representative of the collection based on the cumulatively weighted items and features, wherein each item is represented by a vector that is sum of the vectors of the features comprising the item.

9. The computer program product of claim 8, wherein the first weighting comprises deleting the first identified items from the collection prior to preparing the second vector space.

10. The computer program product of claim 8, wherein:

each item is characterized by a category feature having a category value; and

first identifying items comprises identifying particular items characterized by a vector in the first vector space that is closer to the centroid of vectors associated with at least one other category value than to the centroid of the vectors associated with the category value of the particular item.

11. The computer program product of claim 8, wherein:

each item is characterized by a category feature having a category value, and each category value is represented by the centroid of vectors of the items characterized by the category value; and

first identifying features contributing to first vector space noise comprises;

identifying the N feature vectors closest to the centroid of the item vectors of at least one category value; and

identifying the M feature vectors farther than X feature vectors from the centroid of the item vectors of at least one category value, wherein X is greater than or equal to N; and

first weighting the first identified items and features comprises;

weighting the identified N feature vectors to increase the influence of the N feature vectors on vector spaces representative of the collection, and

weighting the identified M feature vectors to decrease the influence of the M feature vectors on vector spaces representative of the collection.

12. The computer program product of claim 8:

wherein first identifying comprises identifying only a first one of;

items contributing to vector space noise and

features contributing to vector space noise; and

the computer program product further comprises computer-executable program instructions to, after preparing the second vector space representation;

second identify, based on information characteristic of the second vector space, the other one of;

items contributing to second vector space noise and

features contributing to second vector space noise;

second weight the second identified items and features to decrease the influence of the second identified items and features on vector spaces representative of the collection; and

prepare a third vector space representative of the collection based on the cumulatively weighted items and features, wherein each item is represented by a vector that is function of the features comprising the item.

13. The computer program product of claim 8, wherein each prepared vector space is a latent semantic indexed vector space.

14. The computer program product of claim 8:

wherein each item is characterized by a category feature having a category value; and

the computer program product further comprises computer-executable program instructions to;

fold in to the second vector space a vector corresponding an item to be categorized; and

categorize the item to be categorized with the category value of the category feature having a vector in the second vector space closest to the folded-in vector.

15. A system to mitigate noise in vector space representations of item collections, the system comprising:

a storage device; and

a processor communicatively coupled to the storage device, wherein the processor executes application code instructions that are stored in the storage device to cause the system to;

receive a collection of items, each item characterized by a plurality of features;

prepare a first vector space representative of the collection, wherein each item in the first vector space is represented by a vector that is a function of the features characterizing the item;

first identify, based on information characteristic of the first vector space, only one of;

items contributing to first vector space noise, and

features contributing to first vector space noise;

first weight the first identified items and features to decrease the influence of the first identified items and features on vector spaces representative of the collection;

prepare a second vector space representative of the collection based on the first weighted items and features;

second identify, based on information characteristic of the second vector space, the first one of;

items contributing to second vector space noise and

features contributing to second vector space noise;

second weight the second identified items and features to decrease the influence of the second identified items and features on vector spaces representative of the collection; and

prepare a third vector space representative of the collection based on the cumulatively weighted items and features, wherein each item is represented by a vector that is sum of the vectors of the features comprising the item.

16. The system of claim 15, wherein the first weighting comprises deleting the first identified items from the collection prior to preparing the second vector space.

17. The system of claim 15, wherein:

each item is characterized by a category feature having a category value; and

first identifying items comprises identifying particular items characterized by a vector in the first vector space that is closer to the centroid of the item vectors associated with at least one other category value than to the centroid of the item vectors associated with the category value of the particular item.

18. The system of claim 15, wherein:

each item is characterized by a category feature having a category value, and each category value is represented by the centroid of vectors of items characterized by the category value; and

first identifying features contributing to first vector space noise comprises;

identifying the N feature vectors closest to the centroid of the item vectors of at least one category value; and

identifying the M feature vectors farther than X feature vectors from the centroid of the item vectors of at least one category value, wherein X is greater than or equal to N; and

first weighting the first identified items and features comprises;

weighting the identified N feature vectors to increase the influence of the N feature vectors on vector spaces representative of the collection, and

weighting the identified M feature vectors to decrease the influence of the M feature vectors on vector spaces representative of the collection.