An Island-Based Evolutionary Algorithm with a Novel Fitness Approximation Function for the Influence Maximization Problem
Summary
nfluence Maximization (IM) seeks the set of nodes that maximizes expected
influence spread under a diffusion model on a static network. Existing Evolu-
tionary Algorithms (EA) approaches are limited by slow or inaccurate fitness
approximation methods. This thesis contributes (i) ThreeHopSpread, a deter-
ministic multi-hop fitness approximation that captures cascades up to three
hops while remaining computationally efficient, and (ii) an island-based EA
that evolves semi-isolated subpopulations with periodic best-replaces-worst
migration to preserve diversity and encourage exploration of the entire search
space. A node filtering pre-processing step further improves scalability by
restricting the candidate pool of nodes to high-quality options. On real-world
datasets, ThreeHopSpread achieves one of the lowest pairwise ranking error
rates while maintaining a practical runtime, and the island EA under the
Weighted Cascade (WC) model delivers superior or competitive influence
spread in a practical runtime while showing low sensitivity to changes in
the hyperparameter settings. Overall, the method is robust, scalable, and
effective for IM.
