MoELLA-GNN: a hybrid graph neural framework with reinforcement learning path sampling and expert mixtures

Real-world graph data is characterized by heterogeneous structure and evolutionary semantics, thus making learning over complex networks a multi-objective learning problem with a trade-off between accuracy and efficiency. Traditional GNNs lack the intelligent and adaptive mechanisms to model this complexity. We propose MoELLA-GNN, a novel intelligent system that unifies reinforcement-learning path sampling, mixture-of-experts aggregation, and language-informed semantic edge refinement for graph neural networks. Under the constraint of a given computational budget, the model introduces a set of learning-based strategies to adaptively filter multi-hop neighborhoods with high information content; meanwhile, the sparse expert routing module combines a variety of aggregators with complementary properties, including cyclic encoders, transformer encoders, attention-based encoders, domain-averaged encoders, and linguistic semantic encoders. Textual information is involved in the optimization of edge weights to enhance and calibrate the underlying topological relationships. On seven static node classification benchmarks, the method generally achieves optimal or near-optimal performance when compared to a variety of strong baselines, while efficiency metrics such as inference time, GPU memory overhead, and number of model parameters are systematically analyzed. Ablation experiments and sensitivity analysis further reveal the respective contributions of semantic optimization, expert routing, and path strategies. These results indicate that decision-driven sampling and expert mixtures provide a practical route to adaptive and resource-aware learning on complex networks.