Systems and methods for generating dynamic conversational responses based on historical and dynamically updated information

Patent metadata

Field	Value
Publication number	US-2022335222-A1
Application number	US-202117232206-A
Country	US
Kind code	A1
Filing date	Apr 16, 2021
Priority date	Apr 16, 2021
Publication date	Oct 20, 2022
Grant date	—

Abstract

Methods and systems disclosed herein collect user data in real-time and organize them (e.g., using collaborative filtering) into groups (e.g., clusters). The system then determines statistical distributions of observed real-time intents (e.g., based on actual selections made by users) for each group. The system then merges this distribution with existing model predictions (e.g., a model trained on historical training data) to balance between historical and dynamically updated information.

First claim

What is claimed is:

1 . A system for generating dynamic conversational responses based on historical and dynamically updated information, the system comprising:
cloud-based storage circuitry configured to store:
a first machine learning model which performs hybrid collaborative filtering based on similarities in predicted intents and user characteristics between users;
a second machine learning model, wherein the second machine learning model is an unsupervised model; and
a third machine learning model;
cloud-based control circuitry configured to:
receive multi-modal user data in response to a user interacting with a user interface;
generate a first feature input based on the user data;
determine, using the first machine learning model, a first user cluster for the user based on the first feature input, and wherein the first user cluster comprises users having a first actual intent;
determine, using the second machine learning model, a first distribution of probable intents, wherein the first distribution of probable intents is for a subset of users corresponding to the first user cluster, wherein the first distribution of probable intents is based on actual intents of the subset of users during a first time period, and wherein determining the first distribution of probable intents comprises:
generating a second feature input based on the user data; and
determining, using the third machine learning model a second distribution of probable intents based on the second feature input, wherein the second distribution of probable intents is for the user, and wherein the second distribution of probable intents is based on actual intents of users in a plurality of clusters during a second time period;

determining a first probable intent of the user based on a weighted average of the first distribution and the second distribution;
generate for display, on the user interface, a first dynamic conversational response based on the first probable intent of the user;
receive a user selection of the first dynamic conversational response;
determine an actual intent of the user based on the user selection; and
update the third machine learning model based on the actual intent; and
generate for display, on the user interface, a second dynamic conversational response based the third model.

2 . A method for generating dynamic conversational responses based on historical and dynamically updated information, the method comprising:
receiving multi-modal user data in response to a user interacting with a user interface; generating a first feature input based on the user data; determining a first user cluster for the user based on the first feature input, wherein the first user cluster comprises users having a first actual intent; determining a first distribution of probable intents, wherein the first distribution of probable intents is for a subset of users corresponding to the first user cluster, and wherein the first distribution of probable intents is based on actual intents of the subset of users during a first time period; generating a second feature input based on the user data; determining a second distribution of probable intents based on the second feature input, wherein the second distribution of probable intents is for the user, wherein the second distribution of probable intents is based on actual intents of users in a plurality of clusters during a second time period; determining a first probable intent of the user based on a weighted average of the first distribution and the second distribution; and generating for display, on the user interface, a first dynamic conversational response based on the first probable intent of the user.

3 . The method of claim 2 , wherein determining the first user cluster for the user based on the first feature input comprises:
inputting the first feature input into a first model, wherein the first model performs hybrid collaborative filtering based on similarities in predicted intents and user characteristics between users; receiving a first output from the first model; and determining the first user cluster based on the first output.

4 . The method of claim 2 , wherein determining the first distribution of probable intents comprises:
inputting group data for users in the first user cluster into a second model, wherein the second model is an unsupervised model; receiving a second output from the second model; and determining the first distribution based on the second output.

5 . The method of claim 2 , wherein determining the first distribution of probable intents comprises:
determining a number of users in the plurality of clusters; comparing the number to a threshold number; and determining the first distribution of probable intents in response to determining that the number equals or exceeds the threshold number.

6 . The method of claim 2 , wherein determining the first distribution of probable intents comprises:
determining a respective number of users in each of the plurality of clusters; comparing each respective number to a threshold number; and determining the first distribution of probable intents in response to determining that each respective number equals or exceeds the threshold number.

7 . The method of claim 2 , wherein determining the first distribution of probable intents comprises:
determining a number of users in the first user cluster over a period of time; determining a rate of change of the number of users in the first user cluster over the period of time;
comparing the rate of change to a threshold rate of change; and
determining the first distribution of probable intents in response to determining that the rate of change equals or exceeds the threshold rate of change.

8 . The method of claim 2 , wherein determining the second distribution of probable intents based on the second feature input comprises:
inputting the second feature input into a third model, wherein the third model comprises supervised model components; receiving a third output from the third model; and determining the second distribution based on the third output.

9 . The method of claim 8 , further comprising:
receiving a user selection of the first dynamic conversational response; determining an actual intent of the user based on the user selection; and updating the third model based on the actual intent.

10 . The method of claim 2 , wherein determining the first probable intent of the user based on the weighted average of the first distribution and the second distribution comprises:
determining a number of users in the plurality of clusters; and determining a weight to apply in the weighted average based on the number of users in the plurality of clusters.

11 . The method of claim 2 , wherein the multi-modal user data comprises information about the user, cohort assignment, channel information, or a selected intent of the user.

12 . A non-transitory, computer readable medium for generating dynamic conversational responses based on historical and dynamically updated information, comprising instructions that when executed by one or more processors, causes operations comprising:
receiving multi-modal user data in response to a user interacting with a user interface;
generating a first feature input based on the user data;
determining a first user cluster for the user based on the first feature input, wherein the first user cluster comprises users having a first actual intent;
determining a first distribution of probable intents, wherein the first distribution of probable intents is for a subset of users corresponding to the first user cluster, and wherein t

Assignees

• Capital One Services Llc

Inventors

• Le Minh
• Jain Parag

Classifications

• H04L67/566

  Grouping or aggregating service requests, e.g. for unified processing · CPC title

• H04L67/10

  in which an application is distributed across nodes in the network (software deployment G06F8/60; multiprogramming arrangements G06F9/46) · CPC title

• G06F40/35 Primary

  Discourse or dialogue representation · CPC title

• G06N20/00

  Machine learning · CPC title

• H04L67/1097

  for distributed storage of data in networks, e.g. transport arrangements for network file system [NFS], storage area networks [SAN] or network attached storage [NAS] · CPC title