Unveiling the Search Space of Simple Contrastive Graph Clustering with Cartesian Genetic Programming
Maciej Krzywda, Yue Liu, Szymon Łukasik, Amir H. Gandomi · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2025
This paper proposes an enhanced approach to Simple Contrastive Graph Clustering (SCGC) by utilizing Cartesian Genetic Programming (CGP) to evolve neural network architectures. The evolutionary algorithm dynamically optimizes the structure and hyperparameters of SCGC, including the number and size of linear layers, optimizer choice (such as Adam, RMSprop, or SGD), learning rates, weight decay, and loss functions, tailored specifically to the given datasets. Using CGP, we automate both the design and training of SCGC architectures, evolving optimized neural networks with minimal manual intervention. Experimental results conducted across multiple generations on ten benchmark datasets demonstrate that our evolved SCGC consistently achieves superior clustering performance compared to state-of-the-art methods, with notable improvements in accuracy, F1-score, and computational efficiency. The evolutionary process not only optimizes the network topology, but also systematically refines hyperparameters, resulting in robust, highly adaptive, and computationally efficient clustering solutions.