Hierarchical Semantic Risk Minimization for Large-Scale Classification.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Large-scale classification tasks often feature hierarchical label structures, organizing labels from coarse to fine.
• Existing models prioritize prediction accuracy but overlook misclassification risks, crucial in real-world applications.
• Risk in hierarchical classification is measured by the distance between predicted and true labels within the hierarchy.

Purpose of the Study:

• To develop a hierarchical classification model that accounts for and minimizes misclassification risks.
• To introduce a novel risk definition considering both conservative and precipitant factors.
• To enable risk-minimized predictions with flexible granularity in hierarchical classification.

Main Methods:

• Defined classification risk using the semantic hierarchy, incorporating conservative and precipitant risk factors.
• Constructed a balanced conservative/precipitant semantic (BCPS) risk matrix with adjustable weights.
• Modeled classification as a sequential decision-making task using multitask hierarchical learning and deep reinforce multigranularity learning.

Main Results:

• The proposed model effectively reduces classification risk by balancing competing risk factors.
• Achieved superior performance compared to state-of-the-art methods across seven large-scale datasets.
• Demonstrated the capability of obtaining risk-minimized predictions with flexible granularity.

Conclusions:

• The developed model offers a robust solution for hierarchical classification by explicitly managing misclassification risks.
• The BCPS risk matrix and sequential decision-making framework provide a flexible approach to risk-aware classification.
• This work advances hierarchical classification by integrating risk minimization into the prediction process.