Hierarchical Multi-Label Classification System With Dynamic Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques struggle with multi-label classification problems, where data can be classified under multiple non-mutually exclusive labels, leading to complexity in model training and performance evaluation.
Innovation Solution
A computer system is designed with multiple processing instances that share a common code base but run different models, each individually trained for multi-label classification. These models are arranged hierarchically, with outputs from one model serving as input for another, and each model has a unique threshold value for label assignment based on performance analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single model is used for multi-label classification, then the system structure is simple, but the classification accuracy and recall metrics deteriorate due to noise in the data space
Solution Approach 1:
The patent divides the single classification task into multiple specialized models, where each model is trained to classify a specific subset of labels. This segmentation reduces the noise each model needs to handle, improving classification accuracy and recall metrics while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent introduces a hierarchical structure with intermediate processing layers between the input data and final classification outputs. Multiple models process data at different hierarchical levels, with each model serving as an intermediary that refines the classification progressively, thereby improving overall accuracy without requiring a single complex model.
2Measurement precision
If multiple models are used for multi-label classification, then the classification accuracy improves, but the device complexity and training difficulty increase
Solution Approach 1:
The patent employs multiple models that share a common code base and can be trained using a unified framework. Each model is specialized for specific labels but uses the same underlying architecture and training mechanisms, providing universality that reduces the complexity of managing multiple different systems while maintaining high classification accuracy.
Solution Approach 2:
By segmenting the classification task into multiple specialized models with focused label subsets, the patent reduces the complexity each individual model must handle. This segmentation allows for simpler, more efficient training of each model while the collective system achieves high overall accuracy through coordinated operation of the specialized components.
3Measurement precision
If multiple models are used for multi-label classification, then the recall metrics improve, but the training time and computational resources increase
Solution Approach 1:
The patent segments the training process into multiple parallel model training operations, where each model is trained on a specific subset of labels. This segmentation allows for more efficient resource utilization and faster convergence compared to training a single comprehensive model, as each specialized model can be trained more quickly and independently.
Solution Approach 2:
The patent employs different threshold values for different models based on their specific label classification tasks. By optimizing parameters such as classification thresholds for each specialized model, the system achieves improved recall metrics while the modular structure enables efficient parallel training that reduces overall computational time and resource requirements.
Data Source
AI summary
A computer system is provided that is designed to handle multi-label classification. The computer system includes multiple processing instances that are arranged in a hierarchal manner and execute differently trained classification models. The classification task of one processing instance and the executed model therein may rely on the results of classification performed by another processing instance. Each of the models may be associated with a different threshold value that is used to binarize the probability output from the classification model.