Deweighting Formula Corrects Weighted Training Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning classifiers, such as Bayes classifiers, introduce statistical bias when rare classes are over-weighted during training, leading to inaccurate classification probabilities due to the imbalance in training data, causing rare classes to be ignored.
Innovation Solution
A method that involves calculating the mean weight of each class, applying user-defined weights to training data, and using a deweighting formula to correct the probabilities produced by the classifier, ensuring accurate classification decisions by renormalizing probabilities across categories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rare classes are over-weighted during training to improve their classification accuracy, then the classifier pays more attention to rare classes, but statistical bias is introduced causing inaccurate classification probabilities
Solution Approach 1:
The patent applies preliminary action by calculating and storing the mean weight for each class before training occurs. This pre-computed weight information is then used during the classification phase to correct the probabilities, allowing the system to benefit from weighted training while maintaining probability accuracy through pre-prepared correction factors.
Solution Approach 2:
The patent implements feedback by using the classifier's own output probabilities and the pre-computed mean weights to generate corrected probabilities. The correction process feeds back into the classification system, allowing the probabilities to be adjusted based on the weighting scheme used during training, thus compensating for the bias introduced by over-weighting rare classes.
2Ease of operation
If weighted training is applied to balance rare and common classes, then classification decisions on rare classes improve, but the probabilities become biased and inaccurate
Solution Approach 1:
The patent segments the probability correction process into distinct components: calculating mean weights for each class separately, training the classifier with weighted data, and then applying class-specific correction factors to the output probabilities. This segmentation allows the system to maintain accurate classification decisions while correcting probability biases on a per-class basis.
Solution Approach 2:
The patent applies parameter changes by modifying the probability values through a mathematical transformation that incorporates the mean weight parameters. By changing the probability parameters using the formula P(i) = Pw(i) / wi and renormalizing, the system adjusts the output to reflect accurate probabilities while maintaining the benefits of weighted training for classification decisions.
3Reliability
If standard weighting is used to simulate equal class distribution, then rare classes are no longer ignored, but the probabilities for all classes become biased
Solution Approach 1:
The patent introduces an intermediary correction step between the weighted training process and the final probability output. The mean weight calculations serve as an intermediary mechanism that mediates between the weighted training data and the unweighted probability requirements, allowing the system to use weighted training for reliability while producing accurate probabilities through the intermediary correction formula.
Data Source
AI summary
The present invention relates to a method for correcting bias introduced by weighted training in machine learning, the method comprises the following steps: S1. label the number of examples of each class in the weighted data used by the machine learning classifier: sum up the total number of training examples Ni from class in the training set, where i can be any of the class labels; S2. calculate the mean weight of examples in each class: in class i of the training data, sum up the weights wij of each example j, then divide by Ni, and label the mean weight wi of this class i; S3. the features of the examples in the weighted data and their corresponding labels are used for classification and regression with a machine learning algorithm; S4. when the classifier calculates probabilities Pw(i) of class i, and correct them by applying the deweighting formula in the deweighter to get accurate probabilities P(i); S5. use the corrected probabilities P(i) to make a classification decision. The method improves the accuracy of classifiers in assigning probabilities to new data in machine learning applications.