Real-time anomaly determination using integrated probabilistic system

Patent metadata

Field	Value
Publication number	US-11580094-B2
Application number	US-202117331816-A
Country	US
Kind code	B2
Filing date	May 27, 2021
Priority date	May 27, 2021
Publication date	Feb 14, 2023
Grant date	Feb 14, 2023

Abstract

An audio stream is detected during a communication session with a user. Natural language processing on the audio stream is performed to update a set of attributes by supplementing the set of attributes based on attributes derived from the audio stream. A set of filter values is updated based on the updated set of attributes. The updated set of filter values is used to query a set of databases to obtain datasets. A probabilistic program is executed during the communication session by determining a set of probability parameters characterizing a probability of an anomaly occurring based on the datasets and the set of attributes. A determination is made if whether the probability satisfies a threshold. In response to a determination that the probability satisfies the threshold, a record is updated to identify the communication session to indicate that the threshold is satisfied.

First claim

What is claimed is:

1. A system for real-time anomaly detection based on attributes derived from audio streams comprising one or more memory devices storing instructions and one or more processors configured to execute the instructions that, when executed, cause operations comprising:
detecting an audio stream with a user during a communication session;
updating, at each interval of a sequence of intervals of the communication session, a set of attributes related to the user by:
performing natural language processing on the audio stream to supplement the set of user-related attributes with attributes derived from the audio stream during the interval;
updating, during the interval, a set of filtered attributes by selecting attributes of the updated set of user-related attributes based on a set of filter criteria associated with a data filter;
using the updated set of filtered attributes to query a set of databases to obtain anomaly-related datasets related to the set of user-related attributes, wherein each dataset of the anomaly-related datasets is associated with least one filtered attribute of the updated set of filtered attributes;
executing, during the communication session, a probabilistic program by:
using a machine learning model to determine a first probability of an anomaly occurring based on the set of user-related attributes;
determining a second probability of the anomaly occurring by 1) determining a set of probability parameters for a projected probability of anomaly occurrence based on the datasets and 2) determining the second probability based on the probability parameters and the set of filtered attributes; and

in response to a determination that a difference between the first and second probabilities satisfies an anomaly threshold indicating the detection of the anomaly, updating, during the communication session, a record identifying the communication session with an anomaly flag.

2. The system of claim 1 , wherein the set of user-related attributes is a first set of user-related attributes, the instructions further comprising:
retrieving a user record associated with the user,
obtaining a previous set of user-related attributes of the user record, wherein each respective attribute of the previous set of user-related attributes is a same data type as a respective attribute of the first set of user-related attributes;
at each interval of the sequence of intervals, determining a comparison value between an attribute value of the previous set of user-related attributes and an attribute value of the set of user-related attributes, wherein:
updating the set of user-related attributes comprises supplementing the set of user-related attributes with the comparison value;
updating the data filter comprises updating the data filter based on the comparison value; and
executing the probabilistic program comprises determining the second probability based on the comparison value.

3. The system of claim 1 , wherein performing the natural language processing on the audio stream to supplement the set of user-related attributes comprises:
obtaining a set of intent categories associated with different transaction types of database transactions to a set of records associated with the user;
obtaining a set of embedding vectors based on n-grams detected from the audio stream; and
determining an intent category based on the set of embedding vectors, wherein the intent category is associated with a first transaction type, wherein an indicator of the intent category is an attribute of the set of user-related attributes, and wherein updating the set of filtered attributes comprises:
determining whether the first transaction type matches with any transaction types of a recorded set of database transactions associated with the set of records associated with the user; and
in response to a determination that the first transaction type does not match with transaction types of the recorded set of database transactions, updating the set of filtered attributes to include the intent category.

4. The system of claim 1 , wherein updating the attributes comprises:
obtaining a first attribute value for an attribute derived from the audio stream during a first interval of the sequence of intervals;
obtaining a second attribute value for the attribute derived during a second interval of the sequence of intervals;
determining a difference value between the first and second attribute values, wherein:
the updated set of filtered attributes comprises the difference value; and
using the updated set of filtered attributes to query the set of databases comprises using the difference value to query the set of databases.

5. The system of claim 1 , wherein:
using the machine learning model comprises selecting the machine learning model from a set of models based on the set of filtered attributes;
the set of models comprises a neural network model; and
the set of models comprises a regression model.

6. A method comprising:
detecting an audio stream during a communication session with a user;
performing, during the communication session, natural language processing on the audio stream to update a set of attributes based on attributes derived from the audio stream;
updating, at each interval of a set of intervals of the communication session, a set of filter values during the interval based on the set of attributes after the set of attributes is updated;
using the set of filter values after the set of attributes is updated to query a set of databases to obtain datasets associated with the set of filter values;
executing, during the communication session, a probabilistic program by:
determining a set of probability parameters for a projected probability of anomaly occurrence based on the datasets; and
determining a set of probabilities for anomaly occurrence based on the set of probability parameters and the set of attributes; and

in response to a determination that the set of probabilities satisfies a threshold, updating a record identifying the communication session with an anomaly.

7. The method of claim 6 , further comprising:
obtaining a delay duration threshold;
within a first interval, determining a time difference between when an updated set of user-related attributes is provided and when at least one of the set of probabilities is determined;
determining whether the time difference is greater than the delay duration threshold;
in response to a determination that the time difference is greater than the delay duration threshold, reducing a number of queries made to the set of databases to obtain the datasets.

8. The method of claim 6 , further comprising terminating the communication session in response to a determination that the threshold is satisfied.

9. The method of claim 6 , wherein the user is a first user, wherein supplementing the set of attributes comprises:
determining a graphical display of a second user in communication with the first user,
determining whether the set of attributes comprises a first attribute that satisfies a set of prompting criteria;
in response to a determination that the first attribute satisfies the set of prompting criteria:
retrieving a record value of a record identifying the first user; and
displaying a notification comprising the record value on the graphical display.

10. The method of claim 6 , wherein updating the set of filter values comprises:
determining whether a set of filter criteria records indicating the set of filter values was updated;
in response to a determination that the set of filter criteria records was updated, obtain the updated set of filter criteria records, wherein the updated set of filter criteria records co

Assignees

• Capital One Services Llc

Inventors

• Beilis David
• Shpurov Alexey

Classifications

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G10L15/26

  Speech to text systems (G10L15/08 takes precedence) · CPC title

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title