# US 2023/0055980 A1 · Systems and methods for inductive anomaly detection for attributed networks

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

- Number: US 2023/0055980 A1
- Document ID: US20230055980A1
- Status: Published application
- Kind: A1
- Application number: 17/742,236
- Filed: 2022-05-11
- Published: 2023-02-23
- Earliest claimed: 2021-05-11
- Assignees: Arizona Board of Regents on behalf of Arizona State University
- Inventors: Kaize Ding; Huan Liu
- Web page: https://patentgrep.com/patent/US20230055980A1

## CPC

- G06N 3/045: Combinations of networks
- G06N 3/0475: Generative networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Various embodiments of systems and methods for inductive anomaly detection on attributed networks using a graph neural layer to learn anomaly-aware node representations and further employ generative adversarial learning to detect anomalies among new data are disclosed herein.

## Claims

### Claim 1

A system, comprising:
a processor in communication with a memory, the memory including instructions, which, when executed, cause the processor to:
receive, at the processor, a graph indicative of a network that includes a plurality of nodes;
generate, by an encoder network of a graph differentiative network in association with the processor, a set of learned node representations indicative of a plurality of features of the plurality of nodes of the graph, the encoder network being a neural network that includes a plurality of attentional weights that capture one or more feature differences between a node and one or more neighboring nodes of the plurality of nodes; and
generate, by a discriminator network of a generative adversarial network in association with the processor, an output value indicative of a quantitative assessment of a normalcy of the node of the plurality of nodes based on the plurality of features and the one or more feature differences associated with the node of the plurality of nodes, the discriminator network being a neural network.

### Claim 2 (depends on 1)

The system of claim 1, wherein the memory further includes instructions, which, when executed, cause the processor to:
apply an anomaly scoring function to the output value generated by the discriminator network that results in an anomaly score for each respective node of the plurality of nodes.

### Claim 3 (depends on 2)

The system of claim 2, wherein the memory further includes instructions, which, when executed, cause the processor to:
generate a list that includes the plurality of nodes, wherein the list ranks each node of the plurality of nodes based on the anomaly score for each respective node of the plurality of nodes.

### Claim 4 (depends on 1)

The system of claim 1, wherein the encoder network includes a plurality of graph differentiative layers.

### Claim 5 (depends on 4)

The system of claim 4, wherein the memory further includes instructions, which, when executed, cause the processor to:
generate, at each graph differentiative layer of the plurality of graph differentiative layers, a learned node representation of the set of learned node representations for each respective node of the graph, the learned node representation for a node of the plurality of nodes of the graph including a feature difference between the node and a neighboring node of the plurality of nodes in the graph.

### Claim 6 (depends on 1)

The system of claim 1, wherein the graph is indicative of a newly observed network.

### Claim 7 (depends on 1)

The system of claim 1, wherein the memory further includes instructions, which, when executed, cause the processor to:
train the encoder network of the graph differentiative network using a decoder network of the graph differentiative network, the decoder network being a neural network and operable to decode the set of learned node representations and the graph being indicative of a training network.

### Claim 8 (depends on 1)

The system of claim 1, wherein the memory further includes instructions, which, when executed, cause the processor to:
train the discriminator network of the generative adversarial network using a generator network of the generative adversarial network, the generator network being a neural network and operable to generate one or more informative potential anomalies and the graph being indicative of a training network.

### Claim 9 (depends on 8)

The system of claim 8, wherein the memory further includes instructions, which, when executed, cause the processor to:
apply the discriminator network to the one or more informative potential anomalies and one or more learned node representations of the set of learned node representations of the graph;
determine, by the discriminator network, a distribution of normal nodes based on the one or more informative potential anomalies and the set of learned representations of the graph; and
determine, by the discriminator network, a decision boundary that encloses the distribution of normal nodes.

### Claim 10

A system, comprising:
a processor in communication with a memory, the memory including instructions, which, when executed, cause the processor to:
receive, at the processor, a graph indicative of a training network that includes a plurality of nodes;
train an encoder network of a graph differentiative network using a decoder network of the graph differentiative network, the encoder network and the decoder network each being a respective neural network in association with the processor, the encoder network being operable to generate a set of learned node representations of the graph and the decoder network being operable to decode the set of learned node representations; and
train a discriminator network of a generative adversarial network using a generator network of the generative adversarial network, the discriminator network and the generator network each being a respective neural network in association with the processor, the generator network being operable to generate one or more informative potential anomalies and the discriminator network being operable to determine a distribution of normal nodes based on the one or more informative potential anomalies and the set of learned node representations of the graph.

### Claim 11 (depends on 10)

The system of claim 10, wherein the memory further includes instructions, which, when executed, cause the processor to:
determine, by the discriminator network, a decision boundary that encloses the distribution of normal nodes.

### Claim 12 (depends on 10)

The system of claim 10, wherein the memory further includes instructions, which, when executed, cause the processor to:
minimize a first loss that optimizes the generator network on an output of the generator network and an output of the discriminator network; and
minimize a second loss that optimizes the discriminator network on the output of the generator network and the output of the discriminator network, the second loss incorporating the first loss such that a result of the second loss increases when a result of the first loss decreases.

### Claim 13 (depends on 12)

The system of claim 12, wherein the memory further includes instructions, which, when executed, cause the processor to:
jointly optimize the first loss and the second loss.

### Claim 14 (depends on 10)

The system of claim 10, wherein the memory further includes instructions, which, when executed, cause the processor to:
sample, by the generator network, a prior distribution value from a prior distribution; and
generate, by the generator network and using the prior distribution value, the one or more informative potential anomalies.

### Claim 15 (depends on 10)

The system of claim 10, wherein the memory further includes instructions, which, when executed, cause the processor to:
minimize a reconstruction loss between the encoder network and the decoder network.

### Claim 16 (depends on 10)

The system of claim 10, wherein the memory further includes instructions, which, when executed, cause the processor to:
receive, at the processor, a graph indicative of a newly observed network that includes a plurality of nodes;
generate, by the encoder network, a set of learned node representations indicative of a plurality of features of the plurality of nodes of the graph indicative of the newly observed network, the encoder network including a plurality of attentional weights that capture one or more feature differences between a node and one or more neighboring nodes of the plurality of nodes; and
generate, by the discriminator network, an output value indicative of a quantitative assessment of a normalcy of the node of the plurality of nodes based on the plurality of features and the one or more feature differences associated with the node of the plurality of nodes.

### Claim 17

A method, comprising:
receiving, at a processor in association with a memory, a graph indicative of a network that includes a plurality of nodes;
generating, by an encoder network of a graph differentiative network in association with the processor, a set of learned node representations indicative of a plurality of features of the plurality of nodes of the graph, the encoder network being a neural network that includes a plurality of attentional weights that capture one or more feature differences between a node and one or more neighboring nodes of the plurality of nodes; and
generating, by a discriminator network of a generative adversarial network in association with the processor, an output value indicative of a quantitative assessment of a normalcy of the node of the plurality of nodes based on the plurality of features and the one or more feature differences associated with the node of the plurality of nodes, the discriminator network being a neural network.

### Claim 18 (depends on 17)

The method of claim 17, further comprising:
applying, by the processor, an anomaly scoring function to the output value generated by the discriminator network that results in an anomaly score for each respective node of the plurality of nodes.

### Claim 19 (depends on 17)

The method of claim 17, further comprising:
generating, at a graph differentiative layer of a plurality of graph differentiative layers of the encoder network, a learned node representation of the set of learned node representations for each respective node of the graph, the learned node representation for a node of the plurality of nodes of the graph including a feature difference between the node and a neighboring node of the plurality of nodes in the graph.

### Claim 20 (depends on 17)

The method of claim 17, wherein the graph is indicative of a newly observed network.

### Claim 21 (depends on 17)

The method of claim 17, further comprising:
training, by the processor, the graph differentiative network and the generative adversarial network, wherein the graph is indicative of a training network.

### Claim 22 (depends on 21)

The method of claim 21, further comprising:
training the encoder network of the graph differentiative network by the processor and using a decoder network of the graph differentiative network, the decoder network being a neural network and operable to decode the set of learned node representations.

### Claim 23 (depends on 21)

The method of claim 21, further comprising:
training the discriminator network of the generative adversarial network by the processor and using a generator network of the generative adversarial network, the generator network being a neural network and operable to generate one or more informative potential anomalies.

### Claim 24 (depends on 23)

The method of claim 23, further comprising:
applying, by the processor, the discriminator network to the one or more informative potential anomalies and one or more learned node representations of the set of learned node representations of the graph;
determining, by the discriminator network in communication with the processor, a distribution of normal nodes based on the one or more informative potential anomalies and the set of learned representations of the graph; and
determining, by the discriminator network in communication with the processor, a decision boundary that encloses the distribution of normal nodes.
