Predictive Model Input Transformation via Partial Regression Trends

The patent replaces traditional mechanical dimensional reduction methods (stepwise regression, feature selection algorithms) with a causal inference approach based on partial regression plots and causation relationships. This substitution allows variables to be retained based on their causal relevance to the outcome rather than merely their statistical correlation, preventing the removal of high-value variables that have indirect but important causal pathways.

The patent transforms the approach to variable selection by changing from statistical parameters (correlation coefficients, p-values) to causal parameters (causation relationships, partial regression trends). This parameter change enables the identification of variables that may not show strong direct correlation but have important causal relationships through intermediate variables, thus preserving high-value predictors during dimensional reduction.

The patent replaces the mechanical filtering of unstable predictors with a causal analysis system that examines partial regression plots. Variables with inconsistent relationships are not automatically removed but are analyzed for underlying causal structures. The system identifies whether inconsistency arises from confounding factors or intermediate variables, allowing stable retention of predictors that have genuine causal relationships even when direct correlations vary.

The patent introduces causation relationships as intermediary concepts between predictors and outcomes. When a predictor shows inconsistent direct relationship with the outcome, the system searches for intermediate variables that mediate this relationship. The partial regression plot serves as an intermediary analytical tool that decomposes the relationship into direct and indirect effects, allowing the model to maintain stability while preserving predictors with complex causal pathways.

The patent performs preliminary causal analysis using partial regression plots before finalizing the predictive model. This preliminary action involves generating causation relationship maps that identify direct and indirect effects of predictors. By establishing this causal understanding beforehand, the model can make faster predictions without repeatedly analyzing causal relationships, thus reducing computation time while preserving causation understanding.

The patent segments the analysis into distinct phases: (1) exploratory causal analysis using partial regression plots to identify causation relationships, (2) model building with identified causal structures, and (3) prediction execution using the established model. This segmentation allows computationally intensive causal analysis to be performed only once during model development, while subsequent predictions use the pre-established causal framework, thereby reducing overall computation time without losing causation understanding.