Adaptive and Robust DBSCAN With Multi-Agent Reinforcement Learning.

Area of Science:

• Data Mining
• Machine Learning
• Artificial Intelligence

Background:

• Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is effective for arbitrary shapes and noisy data.
• DBSCAN struggles with datasets exhibiting varying density scales, a common real-world challenge.

Purpose of the Study:

• To introduce a novel Adaptive and Robust DBSCAN (AR-DBSCAN) framework utilizing multi-agent reinforcement learning.
• To address the limitations of traditional DBSCAN in handling datasets with diverse density distributions.

Main Methods:

• Data is encoded into a two-level tree, with vertices categorized into density partitions based on information uncertainty.
• Each partition is assigned to an agent for automatic parameter tuning via multi-agent deep reinforcement learning and a Markov decision process.
• A recursive search mechanism optimizes parameter exploration for varying data scales.

Main Results:

• AR-DBSCAN demonstrates significant improvements in clustering accuracy, with up to 144.1% increase in Normalized Mutual Information (NMI) and 175.3% in Adjusted Rand Index (ARI).
• The framework effectively handles datasets with varying density scales and robustly identifies dominant clustering parameters.
• Experiments on artificial and real-world datasets validate the proposed method's performance.

Conclusions:

• AR-DBSCAN successfully overcomes DBSCAN's limitations in varying density environments through adaptive agent allocation and reinforcement learning.
• The proposed method offers a robust and accurate solution for complex clustering tasks, enhancing data mining capabilities.
• AR-DBSCAN provides a scalable and efficient approach for parameter optimization in density-based clustering.