Causal Inference Model Training with Statistical Background Subtraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face challenges in accurately identifying and characterizing the effects of causal factors in demand forecasting due to confounding variables that mask or dilute these effects, leading to inaccurate cause-effect relationships.
Innovation Solution
A model training system that applies deconfounding actions, such as randomized controlled A/B group trials and statistical background subtraction, to separate the effects of causal variables from confounding variables, enabling the direct prediction of individual causal effects on demand quantities like gross margin.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models use historical data and causal factors for demand forecasting, then the model can provide forecasts based on multiple information sources, but confounding variables mask or dilute the effect of causal factors leading to inaccurate cause-effect relationships
Solution Approach 1:
The patent segments the demand signal into multiple components: confounded demand (from historical data), causal demand (from identified causal factors), and residual demand. This segmentation allows the model to separately estimate and combine different demand sources, preventing confounding variables from masking causal effects while still utilizing multiple information sources for forecasting.
Solution Approach 2:
The patent introduces an intermediary causal inference module that processes causal factors through a structured framework. This intermediary layer applies causal inference techniques to isolate the true effect of causal factors from confounding influences, then feeds the purified causal demand signal to the forecasting model, thereby maintaining measurement precision while preserving adaptability.
2Productivity
If machine learning models rely on lagged target time series data and causal factors, then forecasts can be generated, but confounding variables make it difficult to correctly identify and characterize the effect of causal factors
Solution Approach 1:
The patent performs preliminary causal inference analysis before generating forecasts. The system first identifies and characterizes causal factors using causal inference techniques to establish true cause-effect relationships, then uses these validated causal signals as inputs to the forecasting model. This preliminary action ensures reliability of causal identification while maintaining forecast productivity.
Solution Approach 2:
The patent implements a feedback mechanism where the causal inference module continuously refines its understanding of causal relationships based on observed outcomes. The system monitors the impact of causal factors on demand, uses this feedback to adjust causal effect estimates, and incorporates these refined estimates into future forecasts, thereby improving both reliability and productivity over time.
Data Source
AI summary
A system and method are disclosed to generate causal inference machine learning models employing statistical background subtraction. Embodiments include a server comprising a processor and memory. Embodiments receive historical sales data for one or more past time periods and corresponding historical data for one or more causal variables. Embodiments deconfound the cause-effect relationship of historical sales data and historical data on the one or more causal variables. Embodiments define one or more sample weights for statistical background subtraction of the historical data and perform statistical background subtraction on the historical data. Embodiments train a first machine learning model to predict an absolute individual causal effect on a considered demand quantity in relation to the one or more causal variables and one or more sample weights.