# US 2021/0281593 A1 · Systems and methods for machine learning-based digital content clustering, digital content threat detection, and digital content threat remediation in machine learning task-oriented digital threat mitigation platform

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

- Number: US 2021/0281593 A1
- Document ID: US20210281593A1
- Status: Published application
- Kind: A1
- Application number: 17/180,592
- Filed: 2021-02-19
- Published: 2021-09-09
- Earliest claimed: 2020-03-04
- Assignees: Sift Science, Inc.
- Applicants: Sift Science, Inc.
- Inventors: Wei Liu; Jintae Kim; Michael Legore; Yong Fu; Cat Perry; Rachel Mitrano; James Volz; Liz Kao
- Web page: https://patentgrep.com/patent/US20210281593A1

## CPC

- H04L 63/1425: {Traffic logging, e.g. anomaly detection}
- G06F 16/217: {Database tuning (G06F16/2282 takes precedence; database performance monitoring G06F11/3409)}
- G06F 16/2455: Query execution
- G06F 16/285: {Clustering or classification}
- G06F 16/9535: Search customisation based on user profiles and personalisation
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0475: Generative networks
- G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning
- G06N 3/09: Supervised learning
- G06N 3/094: Adversarial learning
- G06N 5/04: Inference or reasoning models
- G06N 20/00: Machine learning
- H04L 63/1416: {Event detection, e.g. attack signature detection}
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A machine learning-based system and method for content clustering and content threat assessment includes generating embedding values for each piece of content of corpora of content data; implementing unsupervised machine learning models that: receive model input comprising the embeddings values of each piece of content of the corpora of content data; and predict distinct clusters of content data based on the embeddings values of the corpora of content data; assessing the distinct clusters of content data; associating metadata with each piece of content defining a member in each of the distinct clusters of content data based on the assessment, wherein the associating the metadata includes attributing to each piece of content within the clusters of content data a classification label of one of digital abuse/digital fraud and not digital abuse/digital fraud; and identifying members or content clusters having digital fraud/digital abuse based on querying the distinct clusters of content data.

## Claims

### Claim 1

A machine learning-based method for content clustering and content threat assessment in a machine learning task-oriented threat mitigation platform, the method comprising:
generating embedding values for each piece of content of one or more corpora of content data;
implementing one or more unsupervised machine learning models that:
(i) receive model input comprising the embeddings values of each piece of content of the one or more corpora of content data; and
(ii) predict a plurality of distinct clusters of content data based on the embeddings values of the one or more corpora of content data;
assessing the plurality of distinct clusters of content data;
associating metadata with each of the plurality of distinct clusters of content data based on the assessment, wherein the associating the metadata includes attributing to each piece of content within the plurality of distinct clusters of content data a classification label of one of (a) an adverse label indicating digital abuse or digital fraud and (b) not digital abuse or not digital fraud;
receiving, via a user interface, a text content query comprising a target piece of online text data associated with one or more online services of one or more subscribers;
querying the plurality of distinct clusters of content data that have the adverse label with the text content query; and
identifying the target piece of online text data with the adverse label indicating digital abuse or digital fraud if one or more of the plurality of distinct clusters of content data that have the adverse label is returned in response to the text content query.

### Claim 2 (depends on 1)

The method according to claim 1, wherein:
the user interface includes a content application programming interface (API) that is searchably connected to the plurality of distinct clusters of content data.

### Claim 3 (depends on 1)

The method according to claim 1, wherein:
the text content query comprises text content observed from an online post or an electronic communication,
the text content is converted to a text embedding value, and
the identifying includes identifying one or more of the plurality of distinct clusters of content data that include pieces of content having the text embedding value.

### Claim 4 (depends on 1)

The method according to claim 1, wherein:
the user interface is a querying interface that includes a tuning interface object that, when adjusted or acted upon by user input, tunes one or more clustering similarity thresholds to increase or decrease a number of members within a target cluster of the plurality of distinct clusters of content data.

### Claim 5 (depends on 4)

The method according to claim 4, further comprising:
querying, via the querying interface, the plurality of distinct clusters of content data based on the text content query;
returning one or more of the plurality of distinct clusters of content data based on the querying; and
increasing or decreasing a number of members within the one or more of the plurality of distinct clusters of content data based on an input to the tuning interface object.

### Claim 6 (depends on 1)

The method according to claim 1, further comprising:
creating a cluster mapping that associates a search grain with at least one cluster of the plurality of distinct clusters of content data.

### Claim 7 (depends on 6)

The method according to claim 6, wherein:
the search grain comprises the target piece of online text data, and
the method further comprising:
using the target piece of online text data to query the plurality of distinct clusters of content data; and
returning, based on the target piece of online text data, one or more clusters of the plurality of distinct clusters of identifiers of the plurality of distinct clusters of content data.

### Claim 8 (depends on 1)

The method according to claim 1, further comprising:
deriving, based on the plurality of distinct clusters of content data, a plurality of distinct clusters of identifiers of a plurality of online users that post online content.

### Claim 9 (depends on 8)

The method according to claim 8, further comprising:
creating a cluster mapping that associates a search grain with at least one cluster of the plurality of distinct clusters of identifiers of the plurality of online users that post online content, wherein the search grain comprises an online user identifier of a user attempting to post online content or posting online content;
using the online user identifier to query the plurality of distinct clusters of identifiers of online users; and
returning, based on the online user identifier, one or more clusters of the plurality of distinct clusters of identifiers of the plurality of online users.

### Claim 10 (depends on 6)

The method according to claim 6, wherein
the search grain comprises an identifier of a subscriber to an online threat service implementing the machine learning-based method,
the method further comprising:
using the identifier of the subscriber to query the plurality of distinct clusters of identifiers of the plurality of online users; and
returning, based on the identifier of the subscriber, one or more cluster members from one or more of the plurality of distinct clusters of identifiers of the plurality of online users.

### Claim 11 (depends on 1)

The method according to claim 1, wherein
the content data relates to text data, communication data, or media data that is posted to a web or Internet-accessible medium, platform, service, system, or channel.

### Claim 12 (depends on 1)

The method according to claim 1, wherein
associating the metadata includes:
associating the classification label, in bulk, to a target cluster of the plurality of distinct clusters of content data, wherein the associating the classification label in bulk causes an association of a single classification label to all members of the target cluster.

### Claim 13 (depends on 1)

The method according to claim 1, wherein:
the identifying includes identifying the one or more of the plurality of distinct clusters of content data based on a query comprising a metadata tag, the metadata tag identifying a classification of the one or more content clusters; and
the method further comprising:
creating a corpus of labeled training data samples for training a target machine learning model based on the one or more content clusters having the associated metadata tag.

### Claim 14 (depends on 1)

The method according to claim 1, wherein
generating the embeddings values includes:
implementing a sentence embeddings or a text embeddings service;
using the sentence embedding or the text embeddings service to generate vector representations for at least a text portion of each piece of content of the one or more corpora of content data.

### Claim 15 (depends on 1)

The method according to claim 1, further comprising:
creating an automated decisioning workflow based on the identifying the one or more of the plurality of distinct clusters of content data having the adverse label indicating digital fraud or digital, wherein the automated decisioning workflow computes, based on machine learning-based scoring of content data, one of a plurality of distinct automated disposal decisions that includes automatically allowing, holding, or disallowing content.

### Claim 16 (depends on 1)

The method according to claim 1, further comprising:
constructing a cluster content data index, that is searchable, by indexing a subset of associated embedding values to each of the plurality of distinct clusters of content data.

### Claim 17 (depends on 16)

The method according to claim 16, further comprising:
searching the cluster content data index based on the text content query to identify one or more clusters or one or more pieces of content from the plurality of distinct clusters of content data that is associated with the text content query.

### Claim 18 (depends on 1)

The method according to claim 1, further comprising:
implementing an automated decisioning workflow for generating blocking instructions or removal instructions of the target piece of online text data of the text content query from an online service of a subscriber if a labeling decision supports the adverse label indicating digital abuse or digital fraud for the text content query.
