Classification Method Using Predictor Segmentation for High Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional classification methods become intractable with large numbers of predictors, often leading to overfitting and poor external cross-validation, as they fail to effectively utilize all available predictors and may include noise, especially in large predictor sets.
Innovation Solution
A method that assigns items to groups based on predictor distributions using statistics, such as confidence intervals, and counts the number of assignments across the predictor space to classify items, rather than relying on discriminant classification functions, allowing for direct classification in large predictor spaces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional discriminant classification methods are used with large numbers of predictors, then the classification functions can be derived, but the methods become intractable and lead to overfitting
Solution Approach 1:
The patent segments the classification problem by treating each predictor independently rather than using joint discriminant functions. Each predictor is evaluated separately through assignment to groups based on its own distribution, avoiding the computational complexity of analyzing all predictors simultaneously in conventional discriminant methods.
Solution Approach 2:
The patent extracts and removes the computationally intractable discriminant classification functions from the analysis. Instead of deriving complex classification functions from all predictors, it extracts simple assignment rules for each predictor based on distribution comparisons, eliminating the mathematical complexity while preserving classification capability.
2Loss of information
If all available predictors are used in conventional methods, then more information is available, but overfitting occurs with excellent training set classification but poor external cross-validation
Solution Approach 1:
The patent segments the predictor set into independent evaluation units. Each predictor is assessed individually through group assignment based on its distribution, allowing all predictors to be utilized without the overfitting that occurs when conventional methods attempt to model relationships among all predictors jointly. This independent segmentation prevents the model from capturing spurious correlations in the training data.
Solution Approach 2:
The patent applies a simplified classification rule (assignment based on distribution) to each predictor individually, using a partial approach rather than attempting to optimize the complete classification function. This partial action on each predictor avoids the overfitting that results from trying to perfectly fit all predictors together, while still utilizing information from all available predictors through their individual contributions.
3Object-affected harmful factors
If a small subset of predictors is selected from a large set, then noise is reduced, but useful information from other predictors is lost
Solution Approach 1:
The patent segments the predictor evaluation into independent individual assessments. Rather than selecting a subset of predictors or weighting them in a discriminant function, each predictor is segmented as an independent unit that contributes to classification through its own assignment rule. This ensures that useful information from all predictors is retained while noise is naturally handled through the distribution-based assignment process for each predictor.
Solution Approach 2:
Each predictor serves itself through its own distribution-based assignment process. The predictors do not require selection or weighting by an external discriminant function; instead, each predictor independently contributes to the classification through its own assignment rule. This self-service approach ensures that all predictors, including potentially useful ones, are utilized without requiring external selection that might exclude valuable information.
Data Source
AI summary
A method of classifying an item to one of a plurality of groups includes providing a plurality of predictors associated with the item. For each predictor, the item is assigned to one of the groups. An assignment number is determined for each group. The item is classified to one of the groups based on the assignment number for each group.