Systems and methods for providing content

Patent metadata

Field	Value
Publication number	US-10817774-B2
Application number	US-201615396303-A
Country	US
Kind code	B2
Filing date	Dec 30, 2016
Priority date	Dec 30, 2016
Publication date	Oct 27, 2020
Grant date	Oct 27, 2020

Abstract

Systems, methods, and non-transitory computer-readable media can receive a first content item having a set of frames. A binary hash code that represents the first content item is generated using at least an aggregation model and an iterative quantization hash model, the binary hash code being determined based at least in part on the set of frames of the first content item. The binary hash code is stored, wherein a similarity between the first content item and a second content item is capable of being measured based at least in part on a comparison of the binary hash code of the first content item and a binary hash code of the second content item.

First claim

What is claimed is:

1. A computer-implemented method comprising:
receiving, by a computing system, a first content item having a set of frames;
generating, by the computing system, a binary hash code that represents the first content item using at least an aggregation model and an iterative quantization hash model, the binary hash code being determined based at least in part on the set of frames of the first content item, the generating further comprising:
transforming, by the computing system, the set of frames of the first content item into a fixed-length feature representation using the aggregation model, the transforming further comprising:
assigning, by the computing system, at least one frame in the set of frames associated with the first content item at least a first weight based on a distance between the at least one frame and a center of a first cluster in a plurality of clusters of reduced-dimension floating point vectors associated with a plurality of frames from a set of training content items and a second weight based on a distance between the at least one frame and a center of a second cluster in the plurality of clusters;
decomposing, by the computing system, the at least one frame into a set of components based at least in part on the first weight and the second weight, wherein the set of components relate one or more portions of the first content item with one or more respective topics; and
storing, by the computing system, the binary hash code of the first content item;
determining, by the computing system, a similarity between the first content item and a second content item, wherein the similarity is measured based at least in part on a comparison of the binary hash code of the first content item and a binary hash code of the second content item.

2. The computer-implemented method of claim 1 , the method further comprising:
training, by the computing system, the aggregation model using a set of training content items, the aggregation model being trained to transform the set of frames of the first content item to a fixed-length representation.

3. The computer-implemented method of claim 2 , wherein training the aggregation model further comprises:
determining, by the computing system, respective high-level floating point vectors for each frame of each content item included in the set of training content items;
transforming, by the computing system, the high-level floating point vectors to corresponding reduced-dimension floating point vectors, the reduced-dimension floating point vectors having fewer dimensions than the high-level floating point vectors; and
clustering, by the computing system, the reduced-dimension floating point vectors into a set of clusters.

4. The computer-implemented method of claim 2 , the method further comprising:
training, by the computing system, the iterative quantization hash model using the set of training content items, the iterative quantization hash model being trained to transform the fixed-length representation of the first content item to the binary hash code.

5. The computer-implemented method of claim 1 , wherein generating the binary hash code that represents the first content item further comprises:
transforming, by the computing system, the fixed-length feature representation to the binary hash code using the iterative quantization hash model.

6. The computer-implemented method of claim 5 , wherein transforming the set of frames of the first content item into the fixed-length feature representation using the aggregation model further comprises:
determining, by the computing system, respective high-level floating point vectors for each frame in the set of frames corresponding to the first content item;
transforming, by the computing system, the high-level floating point vectors to corresponding reduced-dimension floating point vectors, the reduced-dimension floating point vectors having fewer dimensions than the high-level floating point vectors;
pooling, by the computing system, each component in the set of components; and
concatenating, by the computing system, each component in the set of components to produce the fixed-length feature representation.

7. The computer-implemented method of claim 6 , wherein each
frame corresponds to a component, and wherein each component is weighted based on a respective distance between the frame and a cluster center.

8. The computer-implemented method of claim 6 , wherein an average pooling is applied to each of the components.

9. The computer-implemented method of claim 5 , wherein transforming the fixed-length feature representation to the binary hash code using the iterative quantization hash model further comprises:
quantizing, by the computing system, the fixed-length feature representation using the iterative quantization hash model to produce the binary hash code.

10. The computer-implemented method of claim 1 , wherein comparison of the binary hash code of the first content item and the binary hash code of the second content item is performed using one or more bitwise operations.

11. A system comprising:
at least one processor; and
a memory storing instructions that, when executed by the at least one processor, cause the system to perform:
receiving a first content item having a set of frames;
generating a binary hash code that represents the first content item using at least an aggregation model and an iterative quantization hash model, the binary hash code being determined based at least in part on the set of frames of the first content item, the generating further comprising:
transforming the set of frames of the first content item into a fixed-length feature representation using the aggregation model, the transforming further comprising:
assigning at least one frame in the set of frames associated with the first content item at least a first weight based on a distance between the at least one frame and a center of a first cluster in a plurality of clusters of reduced-dimension floating point vectors associated with a plurality of frames from a set of training content items and a second weight based on a distance between the at least one frame and a center of a second cluster in the plurality of clusters;
decomposing, by the computing system, the at least one frame into a set of components based at least in part on the first weight and the second weight, wherein the set of components relate one or more portions of the first content item with one or more respective topics; and
storing the binary hash code of the first content item;
determining a similarity between the first content item and a second content item, wherein the similarity is measured based at least in part on a comparison of the binary hash code of the first content item and a binary hash code of the second content item.

12. The system of claim 11 , wherein the instructions further cause the system to perform:
training the aggregation model using a set of training content items, the aggregation model being trained to transform the set of frames of the first content item to a fixed-length representation.

13. The system of claim 12 , wherein training the aggregation model further causes the system to perform:
determining respective high-level floating point vectors for each frame of each content item included in the set of training content items;
transforming the high-level floating point vectors to corresponding reduced-dimension floating point vectors, the reduced-dimension floating point vectors having fewer dimensions than the high-level floating point vectors; and
clustering the reduced-dimension floating point vectors into a set of clusters.

Assignees

• Facebook Inc

Inventors

• Li Kai
• Yang Fei
• Paluri Balamanohar

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06F16/783 Primary

  using metadata automatically derived from the content · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0454 Primary

  Physics · mapped topic