A Novel Hybrid Machine Learning Framework for Species Influence from Minimal Data

Identifying ecologically influential species is crucial for biodiversity conservation. Yet, classical keystone estimation approaches such as Ecopath-derived KS1, KS2, and KS3 require numerous ecological parameters that are often difficult to obtain and sensitive to uncertainty. To overcome these limitations, we introduce a hybrid machine learning framework that infers species influence using only two widely accessible inputs: diet composition matrices and biomass. The methodology integrates mechanistic descriptors (such as Relative Total Impact) with graph-based topological features (such as PageRank, extended degree centrality) and employs three core learning strategies: Random Forest for supervised prediction, Label Propagation for semi-supervised inference, and GraphSAGE for inductive graph representation learning. These complementary models are combined through an ensemble strategy to generate robust and generalizable species-influence predictions. The framework is evaluated across multiple Ecopath ecosystems containing diet matrices, biomass data, and expert-validated keystoneness rankings. Results demonstrate that the ensemble effectively approximates Ecopath-derived ranking behaviour while requiring far fewer ecological inputs. Importantly, the objective is not to introduce a new keystone metric but to evaluate whether hybrid machine learning models can reliably reproduce expert-derived species rankings. By reducing data requirements and improving cross-ecosystem generalizability, the proposed approach offers a scalable, evidence-driven tool for ecological assessment and conservation planning in data-scarce environments-thereby advancing conservation and supporting SDG-aligned ecosystem management.