Skewed Data Classification via Balanced Subsets and Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for graph classification in database systems fail to effectively handle skewed class distributions, leading to bias towards the majority class and neglect of the minority class, which is critical in identifying rare active classes amidst vast inactive classes.
Innovation Solution
The approach involves constructing multiple classifiers using subsets of training data with balanced positive and negative samples, employing a voting scheme and cascaded feature ensembles to handle skewed distributions and partial feature coverage, and utilizing frequent subgraph mining to select discriminative features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional learning methods are directly applied on skewed data to minimize error rate, then the majority class classification accuracy is improved, but the minority class identification capability deteriorates
Solution Approach 1:
The training data is segmented into multiple balanced subsets through resampling techniques (oversampling minority class, undersampling majority class). Each subset is used to train a separate classifier, ensuring that no single classifier is biased toward the majority class. The ensemble combines predictions from all classifiers to achieve both majority and minority class accuracy.
Solution Approach 2:
The class distribution parameter is changed from skewed (original) to balanced (resampled). By modifying the training data composition to have equal representation of minority and majority classes, the classifiers learn to recognize minority class patterns without being overwhelmed by majority class examples.
2Reliability
If multiple classifiers are constructed on balanced data subsets with different composite features, then the minority class recognition capability is improved, but the system complexity increases
Solution Approach 1:
The complex task of handling skewed data is segmented into multiple simpler subtasks: creating balanced data subsets, training individual classifiers on each subset, and combining their predictions. This segmentation makes the overall system manageable and allows each component to be optimized independently.
Solution Approach 2:
Multiple classifiers with different feature sets are merged into an ensemble system. The individual classifiers, each trained on balanced data with specific composite features, are combined through voting or averaging to produce a final prediction that leverages the strengths of all members, improving minority class recognition while distributing the complexity across modular components.
3Measurement precision
If composite features are selected to differentiate training data with consideration of feature discrimination and class labels, then the classification performance on skewed data is improved, but the feature selection complexity increases
Solution Approach 1:
The feature selection process changes the parameter of feature discrimination by creating composite features that specifically highlight differences between minority and majority classes. This targeted approach to feature construction improves classification performance on skewed data by emphasizing discriminative patterns.
Solution Approach 2:
Composite features are pre-computed and selected before the actual classification task. This preliminary action of feature engineering prepares the data in advance, creating discriminative features that make the subsequent classification easier and more accurate, while consolidating the complexity into a one-time preprocessing step.
Data Source
AI summary
Techniques for classifying structural data with skewed distribution are disclosed. By way of example, a method classifying structural input data comprises a computer system performing the following steps. Multiple classifiers are constructed, wherein each classifier is constructed on a subset of training data, using one or more selected composite features from the subset of training data. A consensus among the multiple classifiers is computed in accordance with a voting scheme such that at least a portion of the structural input data is assigned to a particular class in accordance with the computed consensus. Such techniques for structured data classification are capable of handling skewed class distribution and partial feature coverage issues.