Optimal Surface Selection for Hierarchical Ontology Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In hierarchical classification tasks, machine learning models face challenges due to long tail distributions and hierarchy complexity, where insufficient training data and semantic diversity hinder accurate classification to leaf nodes, necessitating a balance between classification accuracy and coverage.
Innovation Solution
A method is implemented to find an optimal surface in the hierarchical ontology that maximizes accuracy and coverage by selecting a surface that passes through each path from the root to the leaf node, determining adjacent surfaces, and choosing the optimal surface that balances accuracy and coverage, allowing for training of a classifier model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data is manually annotated for hierarchical classification, then classification accuracy can be improved, but the cost and time required for data preparation increases significantly
Solution Approach 1:
The system performs preliminary action by automatically generating synthetic training data through data augmentation techniques before the actual classification task. This pre-processing step creates diverse training samples with varying levels of abstraction from existing annotated data, eliminating the need for extensive manual annotation while maintaining classification accuracy.
Solution Approach 2:
The system changes parameters by varying the level of abstraction in training data samples. It generates synthetic samples at different hierarchical levels (from specific leaf nodes to broader parent nodes), allowing the model to learn both fine-grained and coarse-grained classifications without requiring manual annotation for each level.
2Adaptability or versatility
If the ontology hierarchy is made more detailed with more leaf nodes, then coverage of specific classes improves, but classification accuracy decreases due to insufficient training data for each class
Solution Approach 1:
The system segments the training task by dividing it into multiple levels of abstraction. Instead of treating all leaf nodes uniformly, it segments the hierarchy into different tiers and generates synthetic training data at each tier, allowing the model to progressively learn from general to specific classes.
Solution Approach 2:
The system applies partial action by generating synthetic training data selectively for underrepresented classes and hierarchical levels. It focuses computational resources on generating data where it is most needed (classes with insufficient samples) rather than uniformly distributing effort across all classes.
3Adaptability or versatility
If a classifier model is trained to classify to all leaf nodes, then coverage is maximized, but accuracy suffers due to long tail distribution of training data
Solution Approach 1:
The system introduces dynamics by making the classification target flexible rather than fixed. It dynamically adjusts which hierarchical level serves as the classification target based on data availability and model performance, allowing the system to adapt between classifying to leaf nodes when data is sufficient and classifying to parent nodes when data is scarce.
Solution Approach 2:
The system uses intermediate hierarchical levels as mediators between the root and leaf nodes. For classes with insufficient training data, it introduces parent nodes as intermediate classification targets, allowing the model to achieve reasonable accuracy at intermediate levels while still providing a pathway to more specific classifications when possible.
Data Source
AI summary
A mechanism is provided in a data processing system having a processor and a memory. The memory comprises instructions which are executed by the processor to cause the processor to implement a training system for finding an optimal surface for hierarchical classification task on an ontology. The training system receives a training data set and a hierarchical ontology data structure. A surface finding component executing within the training system selects a surface that passes through each path from a root to a leaf node in the hierarchical ontology data structure. The surface finding component determines a plurality of adjacent surfaces that differ from the selected component by one node. The surface finding component selects an optimal surface, based on the selected surface and the plurality of adjacent surfaces, that maximizes accuracy and coverage. The training system trains a classifier model for a cognitive system using the optimal surface and the training data set.