# US 12,182,681 B2 · Information processing device, information processing method, and storage medium

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,182,681 B2
- Document ID: US12182681B2
- Status: Granted
- Kind: B2
- Application number: 17/098,508
- Filed: 2020-11-16
- Published: 2021-03-04
- Granted: 2024-12-31
- Earliest claimed: 2018-05-23
- Assignees: Kabushiki Kaisha Toshiba; Toshiba Digital Solutions Corporation
- Applicants: Kabushiki Kaisha Toshiba; Toshiba Digital Solutions Corporation
- Inventors: Hidemasa Itou; Yukio Kamatani; Katsuyuki Hanai
- Web page: https://patentgrep.com/patent/US12182681B2

## CPC

- G06N 3/04: Architecture, e.g. interconnection topology
- G06F 17/16: Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)}
- G06F 18/24133: {Distances to prototypes}
- G06F 18/29: {Graphical models, e.g. Bayesian networks}
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/063: using electronic means
- G06N 3/08: Learning methods
- G06N 5/022: Knowledge engineering; Knowledge acquisition
- G06N 3/048: Activation functions
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

An information processing device of embodiments includes a data acquirer and a network processor. The data acquirer is configured to acquire graph structure data that includes a plurality of real nodes and one or more real edges connecting two of the plurality of real nodes. The network processor is configured to execute processing of propagating a feature amount of a k−1th layer of each of a plurality of assumed nodes that include the plurality of real nodes and the one or more real edges at least to a feature amount of a kth layer of another assumed node in a connection relationship with each of the assumed nodes in a neural network on the basis of the graph structure data acquired by the data acquirer. k is a natural number equal to or more than 1.

## Claims

### Claim 1

An information processing device comprising:
processing circuitry configured to
acquire graph structure data that includes a plurality of real nodes and one or more real edges connecting two of the plurality of real nodes; and
execute processing of setting a neural network that propagates a feature amount of each of a plurality of assumed nodes, the plurality of assumed nodes including the plurality of real nodes and the one or more real edges, that are distinguishable from each other, and propagating a feature amount of a k-1th layer of each of the plurality of assumed nodes at least to a feature amount of a kth layer of another assumed node in a connection relationship with each of the assumed nodes in the neural network on the basis of the acquired graph structure data, k being a natural number equal to or more than 1,
wherein the processing circuitry is configured to set at least part of the coefficient when the feature amount is propagated using a rule based on a graph attention network, and
wherein the processing circuitry is configured to set a propagation matrix to be applied to each of the feature amount of an assumed node of a propagation source and the feature amount of an assumed node of a propagation destination according to the type of the assumed node of a propagation source.

### Claim 2 (depends on 1)

The information processing device according to claim 1,
wherein the processing circuitry is configured to propagates the feature amount of the k-1th layer of each of the plurality of assumed nodes to a feature amount of a kth layer of each of the assumed nodes itself.

### Claim 3 (depends on 1)

The information processing device according to claim 1,
wherein the processing circuitry is configured to set types of the assumed nodes from a plurality of types, and make a rule for setting a coefficient when the feature amount is propagated different depending on the set types of the assumed nodes.

### Claim 4

An information processing method comprising:
by a computer,
acquiring graph structure data that includes a plurality of real nodes and one or more real edges connecting two of the plurality of real nodes; and
executing processing of setting a neural network that propagates a feature amount of each of a plurality of assumed nodes, the plurality of assumed nodes including the plurality of real nodes and the one or more real edges, that are distinguishable from each other, and propagating a feature amount of a k-1th layer of each of the plurality of assumed nodes at least to a feature amount of a kth layer of another assumed node in a connection relationship with each of the assumed nodes in the neural network on the basis of the acquired graph structure data, k being a natural number equal to or more than 1,
wherein the information processing method comprises:
setting at least part of the coefficient when the feature amount is propagated using a rule based on a graph attention network; and
setting a propagation matrix to be applied to each of the feature amount of an assumed node of a propagation source and the feature amount of an assumed node of a propagation destination according to the type of the assumed node of a propagation source.

### Claim 5

A non-transitory computer-readable storage medium storing program causing a computer to perform:
acquiring graph structure data that includes a plurality of real nodes and one or more real edges connecting two of the plurality of real nodes; and
executing processing of setting a neural network that propagates a feature amount of each of a plurality of assumed nodes, the plurality of assumed nodes including the plurality of real nodes and the one or more real edges, that are distinguishable from each other, and propagating a feature amount of a k-1th layer of each of the plurality of assumed nodes at least to a feature amount of a kth layer of another assumed node in a connection relationship with each of the assumed nodes in the neural network on the basis of the acquired graph structure data, k being a natural number equal to or more than 1,
wherein the program causes the computer to perform:
setting at least part of the coefficient when the feature amount is propagated using a rule based on a graph attention network; and
setting a propagation matrix to be applied to each of the feature amount of an assumed node of a propagation source and the feature amount of an assumed node of a propagation destination according to the type of the assumed node of a propagation source.