Transfer-Learning-Based Gaussian Mixture Model for Distributed Clustering.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Distributed clustering using Gaussian mixture models (GMM) is effective in peer-to-peer (P2P) networks.
• Existing methods require many iterations and high communication overhead.
• GMM parameter updates lack a closed form, leading to imprecise clustering accuracy.

Purpose of the Study:

• To develop a general transfer distributed GMM clustering framework.
• To enhance clustering performance and accelerate convergence in P2P networks.
• To address limitations of existing distributed GMM clustering algorithms.

Main Methods:

• Utilized transfer learning, treating each node as both source and target domains.
• Developed a transfer distributed expectation-maximization (EM) algorithm with a fixed learning rate.
• Introduced an improved version with an adaptive transfer learning strategy to adjust the learning rate automatically.
• Extended a representative GMM method (entropy-type classification maximum-likelihood) to its transfer distributed form.

Main Results:

• The proposed framework significantly promotes clustering performance.
• Accelerated clustering convergence was observed compared to existing approaches.
• Experimental results verified the effectiveness of the presented transfer distributed algorithms.

Conclusions:

• The transfer distributed GMM clustering framework effectively improves clustering accuracy and convergence speed.
• The adaptive learning rate strategy ensures stable clustering accuracy.
• The framework is extensible and demonstrates superior performance over traditional methods.