Resilient Data Classification via Ensemble SVMs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification systems, such as SVMs, are vulnerable to errors in training data, which can lead to incorrect classification performance due to malicious or unreliable data sources, especially when training data is sourced from multiple potentially erroneous providers.
Innovation Solution
A classification system that creates multiple SVMs based on small, randomly selected subsets of training data and combines their results using methods like simple or weighted majority votes to enhance resilience against errors in the training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single SVM is trained on the entire training data set, then the classification model can be simple and fast to train, but it becomes vulnerable to errors in the training data and may be easily manipulated by malicious entities
Solution Approach 1:
The patent divides the training data into multiple small subsets and trains separate SVM classifiers on each subset. This segmentation approach prevents any single malicious or erroneous data point from dominating the entire training process, as each subset is small enough that errors cannot significantly alter the hyper-surface. The multiple classifiers work together to provide robust classification despite individual data errors.
2Reliability
If multiple classifiers are created based on small subsets of training data, then resilience against errors in training data is improved, but the complexity of the classification system increases
Solution Approach 1:
The patent combines multiple SVM classifiers into a single combined classifier by merging their Lagrange multipliers and bias terms. This merging process integrates the knowledge from all individual classifiers trained on different subsets, creating a unified model that maintains the resilience benefits of multiple classifiers while simplifying the system structure. The combined classifier can be applied directly without managing multiple separate models.
3Measurement precision
If a large portion of training data is used to train a single SVM, then training accuracy can be maximized, but the system becomes more susceptible to correlated errors from malicious or unreliable data sources
Solution Approach 1:
The patent segments the large training data set into many small subsets, training one SVM on each subset. This segmentation ensures that even if a large portion of the total data is used for training, no single subset contains enough erroneous data to significantly alter the hyper-surface. The distributed training approach across multiple subsets provides inherent protection against correlated errors while maintaining high training accuracy through the aggregation of results from all subsets.
Data Source
AI summary
A classification system is described for resilient classification of data. In various embodiments, the classification system divides a full set of the training data into a positive subset and a negative subset. The positive subset contains all training data with a positive classification value and the negative subset contains all training data with a negative classification value. The classification system constructs several subsets of the training data wherein each subset contains data randomly selected from both the positive subset and the negative subset. The classification system then creates at least two classifiers based on each of the randomly created subsets of the training data.