A Novel Active Learning Regression Framework for Balancing the Exploration-Exploitation Trade-Off.

Area of Science:

• Machine Learning
• Optimization Theory

Background:

• Active learning is crucial for cost-effective data acquisition in real-world applications.
• Existing active learning methods primarily focus on improving model accuracy, neglecting other domain-specific objectives.

Purpose of the Study:

• To develop a novel active learning framework for general optimization problems, addressing the exploration-exploitation trade-off.
• To investigate and compare pool-based and query synthesis active learning approaches.
• To apply the framework to regression tasks, specifically learning price-demand functions for optimal pricing strategies.

Main Methods:

• A comprehensive active learning framework incorporating exploration-based, exploitation-based, and balancing strategies.
• Investigation of regression tasks, which are less explored in active learning compared to classification.
• Comparative analysis of pool-based and query synthesis querying approaches.

Main Results:

• The proposed framework demonstrates significant performance in optimizing general problems beyond accuracy.
• Effective balancing of exploration and exploitation strategies was achieved.
• Successful application and performance validation on the price-demand function learning problem.

Conclusions:

• The novel active learning framework offers a versatile approach for various optimization problems, especially in regression.
• The framework provides effective strategies for managing the exploration-exploitation dilemma in data acquisition.
• The proposed methods show superior performance compared to baseline approaches in experimental evaluations.