A Distributed Learning Method for ℓ 1 -Regularized Kernel Machine over Wireless Sensor Networks

Area of Science:

• Machine Learning
• Wireless Sensor Networks
• Distributed Systems

Background:

• Centralized learning in wireless sensor networks (WSNs) incurs high communication and energy costs due to transmitting raw data to a central node.
• Existing methods struggle to balance model accuracy with resource efficiency in WSNs.

Purpose of the Study:

• To propose a novel distributed learning algorithm for ℓ 1 -regularized kernel minimum mean squared error (KMSE) machines.
• To reduce communication costs and energy consumption in WSNs through in-network processing and sparse model transmission.

Main Methods:

• Developed a distributed learning algorithm for ℓ 1 -regularized KMSE machines.
• Implemented in-network processing and inter-node collaboration for sparse model exchange between single-hop neighbors.
• Evaluated performance using prediction accuracy, model sparsity, communication cost, and iteration count on synthetic and real datasets.

Main Results:

• The proposed algorithm achieves prediction accuracy comparable to batch learning methods.
• Demonstrated significant improvements in model sparsity and communication cost reduction.
• Showcased faster convergence with fewer iterations compared to existing approaches.
• Validated advantages through experiments on a WSN test platform.

Conclusions:

• The novel distributed learning algorithm effectively addresses the communication and energy challenges in WSNs.
• The algorithm offers a superior trade-off between model accuracy, sparsity, and communication efficiency.
• This approach is well-suited for resource-constrained environments like wireless sensor networks.