GP-DPC: Genetic Programming-Driven Nonlinear Feature Projection for Scalable Community Detection in Attributed Networks

Attributed networks are widely used to model engineering systems in which both relational structure and node-specific information are important, such as social platforms, transportation networks, and knowledge graphs. Community detection in such networks requires methods that can jointly exploit graph topology and heterogeneous node attributes while remaining computationally feasible. To address this problem, this paper proposes GP-DPC, a genetic programming-driven framework for projectionguided community detection in attributed networks. Instead of evolving pairwise similarity functions, GP-DPC uses a multi-tree GP representation to learn three nodelevel nonlinear projection functions, which map attributed nodes into a compact lowdimensional space. Fast density peaks clustering is then performed in the learned space through an approximate kNN strategy, avoiding full pairwise distance computation. A community-aware fitness function combines structural modularity and projection-space separability to guide the evolutionary search. Experiments cover synthetic Handl and LFR benchmarks and five real-world attributed networks. Direct comparisons on the Handl 10c500d and 24c1800d benchmarks and on Cora, Citeseer, and ACM show competitive detection accuracy, while runtime analyses indicate practical inference efficiency. These results support GP-DPC as a promising framework for engineeringoriented attributed network community detection.