Regression Model for Positional Data Using Segmented Weight Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning techniques, such as geographically and temporally weighted regression and piecewise-constant conditional intensity models, struggle to efficiently predict outcomes for data records with positional values, particularly in complex data structures like electronic medical records and speech texts, due to limitations in handling positional information for each explanatory variable and increasing model complexity.
Innovation Solution
A computer-implemented method using a machine learning model with a network structure that incorporates positional parameters and weight parameters to evaluate the influence of labels on a target outcome, allowing for efficient prediction of outcomes by training the model with reduced training data and providing insights into the effective range of influence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If geographically and temporally weighted regression (GTWR) is used to analyze positional information, then the model can consider the location of one sample, but it cannot adapt to situations where positional information is given for each explanatory variable in one sample
Solution Approach 1:
The patent segments the positional information processing by introducing separate positional parameters (spatial and temporal) for each explanatory variable. This allows the model to handle different data structures (electronic medical records, speech texts, etc.) with positional information at the variable level rather than requiring complete model restructuring, thus improving adaptability while controlling complexity.
Solution Approach 2:
The patent creates a universal regression model framework that can handle multiple types of positional data (spatial coordinates, timestamps, sequence positions) through a unified set of positional parameters and weight functions. This multi-functional approach allows the same model structure to adapt to various data structures including electronic medical records and speech texts without requiring separate specialized models.
2Ease of operation
If the Piecewise-Constant Conditional Intensity Model is used to learn temporal dependencies, then a closed-form Bayesian approach can be applied, but the contribution from occurrence of certain events on occurrence probability of other events is incorporated into the constant part, making it impossible to incorporate hypotheses that the contribution decays over time
Solution Approach 1:
The patent introduces dynamic temporal weight functions that can model time-decaying contributions of events. Instead of using constant intensity functions, the model employs temporal weight functions w(t, t_j) that vary with time, allowing the contribution of past events to decay over time. This dynamic approach maintains computational tractability while enabling the model to capture temporal decay patterns in event influences.
Solution Approach 2:
The patent changes the parameter structure from constant intensity values to time-varying weight functions with learnable parameters. By parameterizing the temporal weight functions (e.g., using exponential decay parameters or other temporal patterns), the model can flexibly represent time-decaying contributions while maintaining a manageable number of parameters through sharing across event types.
3Quantity of substance
If the Piecewise-Constant Conditional Intensity Model is used, then some parameters related to uncorrelated attributes can be removed, but the number of parameters of the model increases basically in the order of the square of the number of attributes
Solution Approach 1:
The patent merges the handling of multiple attributes by introducing a unified weight function structure that processes all explanatory variables through shared positional parameter sets. Instead of maintaining separate parameter matrices for each attribute pair (leading to O(n²) parameters), the model combines them into a unified framework where positional parameters are shared across attributes, reducing the parameter count from quadratic to linear or near-linear scaling with the number of attributes.
Solution Approach 2:
The patent transforms the parameter structure from a dense O(n²) parameter matrix to a sparse, structured parameterization using positional parameters (spatial coordinates, temporal values) and weight functions. This parameter change exploits the underlying structure of positional relationships to reduce the number of independent parameters from quadratic to a much smaller set that scales more favorably with the number of attributes.
4Productivity
If conventional machine learning techniques are used for data records with positional values, then basic prediction can be performed, but the prediction efficiency and accuracy for complex data structures like electronic medical records and speech texts is insufficient
Solution Approach 1:
The patent applies local quality by introducing position-specific weight functions that adaptively weight the contribution of each explanatory variable based on its positional relationship to the target. Instead of uniform weighting, the model assigns different weights to variables based on their spatial and temporal distances, allowing the prediction to capture local patterns and relationships specific to each data record's positional structure, thereby improving both efficiency and accuracy for complex data types.
Data Source
AI summary
A method, a system, and a computer program product for predicting an outcome expected for a particular positional value is provided. In the method, an input set of data records, each having a label and a positional value, and a target positional value are obtained. The label of each data record is one in a label set. A learning model that includes an output layer, an input layer corresponding to the label set and a network structure provided therebetween is read. In the learning model, the network structure has a plurality of functions trained so as to evaluate influence from each label in the label set depending on a relationship between the target positional value and a representative positional value associated with the label in the label set. A target outcome is estimated for the target positional value from the input set using the learning model.