Adaptive Data Discretization for Multi-Feature Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional discretization methods in machine learning fail to consider the distribution of objective variable values in combination with multiple explanatory variables, leading to suboptimal cutoff points and reduced model accuracy.
Innovation Solution
A machine learning apparatus that performs discretization based on a predetermined criterion, selects important combinations of data items, and adjusts the discretization criterion based on the execution results of training processing, considering the distribution of training data in a multi-dimensional space.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If discretization is performed using conventional methods considering only single explanatory variables, then the processing is simple and fast, but the model accuracy is reduced due to suboptimal cutoff points
Solution Approach 1:
The patent segments the discretization process into two distinct phases: (1) initial discretization based on single explanatory variables to obtain baseline feature importance, and (2) secondary discretization focused only on combinations of explanatory variables identified as important in the first phase. This segmentation allows the system to achieve high model accuracy through multi-variable analysis while avoiding the computational complexity of evaluating all possible variable combinations from the outset.
Solution Approach 2:
Instead of performing exhaustive discretization analysis on all possible combinations of explanatory variables (which would be excessively complex), the patent applies partial action by focusing only on combinations involving explanatory variables that demonstrated importance in the initial single-variable discretization phase. This selective approach achieves improved accuracy without the full computational burden of complete combinatorial analysis.
2Measurement precision
If discretization considers all combinations of explanatory variables to improve accuracy, then model precision improves, but computational load increases significantly
Solution Approach 1:
The patent performs preliminary discretization analysis on individual explanatory variables first to identify which variables have important relationships with the objective variable. This preliminary action creates a filtered set of candidate variables that are then used in the second phase of combination-based discretization, significantly reducing the computational load compared to analyzing all possible combinations from the start.
Solution Approach 2:
The discretization process is divided into two sequential stages: Stage 1 performs discretization on single explanatory variables to establish baseline importance metrics, and Stage 2 performs discretization on combinations only for variables identified as important in Stage 1. This segmentation transforms an O(n^k) complexity problem into a much more manageable two-phase process with significantly reduced computational requirements.
3Adaptability or versatility
If discretization uses fixed criteria without adjustment, then the processing is straightforward, but the model fails to adapt to important feature combinations
Solution Approach 1:
The patent implements a feedback mechanism where the results of the first discretization phase (feature importance metrics) directly influence the second discretization phase. The system uses the acquired knowledge about which explanatory variables are important to adjust and refine the discretization criteria for combination features, creating an adaptive two-stage process that improves model versatility while maintaining manageable complexity through iterative refinement.
Solution Approach 2:
The discretization methodology transitions from a static, single-criteria approach to a dynamic two-stage process. The discretization criteria are not fixed but are adjusted based on feedback from the first phase: variables identified as important trigger more sophisticated combination-based discretization, while less important variables retain simpler discretization treatment. This dynamic adaptation enhances model versatility without uniformly increasing complexity across all features.
Data Source
AI summary
With respect to training data records in which combinations of data item values of data items are individually associated with label information, the data item values are converted based on a criterion per data item into discretized data values. Training processing for training a model that receives the discretized data values as input and performs determination about the label information is performed by using training data records obtained by the conversion. From an execution result of the training processing, feature data records, each of which differently indicates a combination of two or more data items for the determination among the data items, and index values, which indicate importance levels of the feature data records respectively, are acquired. The criterion for the discretization of the data item values is changed based on at least one of the feature data records having been selected based on the index values.