GENERATION OF WEIGHTS FOR CAUSAL INFERENCES

Number: 20240054343
Published: 2024-02-15
Filed: 2023-07-24
Assignee: Disney Enterprises, Inc.
Inventors: Gallusser; Fabian et al.
CPC: G06N3/08; G06N20/00
Verdict: Set aside causal-inference ML for recommendation, business
Source: Google Patents · FreePatentsOnline

Abstract

In some embodiments, a method receives input data to calculate an effect of a variable on a group for a plurality of methods. Methods in the plurality of methods calculate the effect of the variable for the input data using different logic. A plurality of sub-weights for methods in the plurality of methods are generated. The sub-weights are generated based on a balance metric, a dissimilarity metric, and a reliability metric. The method combines the plurality of sub-weights for methods in the plurality of methods to generate a final weight for the methods. The respective final weight is applied to an intermediate result from a respective method in the plurality of methods to generate a weighted intermediate result for the method. The method combines weighted intermediate results for the plurality of methods to generate a final result for the effect of the variable.

Background

BACKGROUND

A system may analyze data for a company. For example, a company may want to assess the causal inference of a treatment on a group, which may be a group of members (e.g., users, devices, households, etc.). The system may be modeled to analyze data to determine an answer for the following question of “What is the impact of X on Y?”. The variable “X” may be the treatment and the variable “Y” may be a metric for the group. The metric may be churn rate for a household, engagement for a user account, latency rate on a device, etc. In some examples, the specific questions may be answered of: “What is the impact of streaming a title of movie name Z on the long-term value of users?”, “What is the impact of installing an application on mobile devices for future retention of users?”, “What is the value of making a title available in another language on engagement?”, etc.

An issue in determining answers to these questions may be the presence of bias, such as self-selection bias, and confounding. Self-selection bias may be where individuals may have the ability to choose whether to participate in the groups, and these individuals may have their own preferences, which may result in bias that may not be truly representative of the entire population. Confounding may be where the relationship between X and Y may be distorted by another variable. For example, the group exposed to the treatment X is not necessarily comparable to the group that is not exposed, so a direc

Claims

1. A system comprising: a plurality of nodes configured to receive input data to calculate an effect of a variable on a group for a plurality of methods, wherein methods in the plurality of methods calculate the effect of the variable for the input data using different logic; an intermediate weight computation system configured to: generate a plurality of sub-weights for methods in the plurality of methods, wherein the sub-weights are generated based on a balance metric, a dissimilarity metric, and a reliability metric; combine the plurality of sub-weights for methods in the plurality of methods to generate a final weight for the methods; and apply the respective final weight to an intermediate result from a respective method in the plurality of methods to generate a weighted intermediate result for the method; and an integration system configured to combine weighted intermediate results for the plurality of methods to generate a final result for the effect of the variable. || 17. A method comprising: receiving, by a computing device, input data to calculate an effect of a variable on a group for a plurality of methods, wherein methods in the plurality of methods calculate the effect of the variable for the input data using different logic; generating, by the computing device, a plurality of sub-weights for methods in the plurality of methods, wherein the sub-weights are generated based on a balance metric, a dissimilarity metric, and a reliability metric; combining, by the computing device, the plurality of sub-weights for methods in the plurality of methods to generate a final weight for the methods; applying, by the computing device, the respective final weight to an intermediate result from a respective method in the plurality of methods to generate a weighted intermediate result for the method; and combining, by the computing device, weighted intermediate results for the plurality of methods to generate a final result for the effect of the variable. || 20. A non-transitory computer-readable storage medium comprising instructions for controlling the one or more computer processors to be operable for: receiving input data to calculate an effect of a variable on a group for a plurality of methods, wherein methods in the plurality of methods calculate the effect of the variable for the input data using different logic; generating a plurality of sub-weights for methods in the plurality of methods, wherein the sub-weights are generated based on a balance metric, a dissimilarity metric, and a reliability metric; combining the plurality of sub-weights for methods in the plurality of methods to generate a final weight for the methods; applying the respective final weight to an intermediate result from a respective method in the plurality of methods to generate a weighted intermediate result for the method; and combining weighted intermediate results for the plurality of methods to generate a final result for the effect of the variable.