Bandit-Based Task Assignment for Heterogeneous Crowdsourcing.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Crowdsourcing

Background:

• Crowdsourcing aims to collect reliable data labels efficiently within budget constraints.
• Worker reliability varies across different task types, posing a challenge for optimal task assignment.
• Heterogeneous crowdsourcing environments require methods that account for task-specific worker expertise.

Purpose of the Study:

• To propose a novel contextual bandit formulation for task assignment in heterogeneous crowdsourcing.
• To address the exploration-exploitation trade-off in selecting workers for diverse tasks.
• To enhance the collection of reliable labels under budget limitations.

Main Methods:

• A contextual bandit model was developed for dynamic task assignment.
• The model considers task characteristics and worker-specific reliability.
• Theoretical analysis of regret bounds was performed to evaluate the method's efficiency.

Main Results:

• The proposed contextual bandit approach effectively manages worker selection in heterogeneous crowdsourcing.
• Theoretical analysis provides bounds on the method's performance.
• Experimental results demonstrate the practical utility and effectiveness of the approach.

Conclusions:

• The developed contextual bandit formulation offers an effective solution for task assignment in heterogeneous crowdsourcing.
• The method optimizes the balance between exploring new worker capabilities and exploiting known worker performance.
• This approach enhances the efficiency and reliability of crowdsourced data collection.