Heuristic optimization in classification atoms in molecules using GCN via uniform simulated annealing

Area of Science:

• Machine Learning
• Deep Learning
• Artificial Intelligence

Background:

• Graph Neural Networks (GNNs) excel at processing graph-structured data, preserving spatial dependencies for improved classification accuracy.
• Training GNNs is computationally intensive and challenging, necessitating advanced optimization techniques.
• Metaheuristic algorithms like genetic algorithms and particle swarm optimization have been applied to neural network hyperparameter tuning.

Purpose of the Study:

• To propose a novel metaheuristic optimization algorithm for Graph Convolutional Networks (GCNs).
• To enhance the training efficiency and performance of GCNs using a hybrid optimization approach.
• To evaluate the proposed method against state-of-the-art optimization techniques on benchmark datasets.

Main Methods:

• Development of a hybrid optimization algorithm combining Simulated Annealing with Uniform distribution and gradient optimizers for GCN weight optimization.
• Implementation of the proposed method and comparison with existing gradient-based and heuristic optimization models.
• Experimental validation on the QM7 dataset, utilizing both balanced and imbalanced data splits.

Main Results:

• The proposed Simulated Annealing-based hybrid optimization method demonstrated superior performance compared to standalone gradient and heuristic optimization models.
• Achieved lower loss function values across tested configurations.
• Exhibited higher accuracy on the balanced dataset and improved AUC (macro) values on the imbalanced dataset.

Conclusions:

• The novel hybrid optimization strategy significantly enhances GCN performance, offering a more effective training approach.
• The method provides a robust solution for optimizing GCNs, particularly in scenarios with imbalanced data.
• This research contributes a valuable tool for advancing deep learning applications utilizing graph-structured data.