The role of graph information in the context of outcome-oriented prediction: an empirical analysis framework

Abstract Graph-structured information is increasingly exploited in outcome-oriented prediction tasks, yet its exact contribution remains empirically unquantified. To address this gap, a controlled evaluation framework is introduced that contrasts four traditional machine-learning pipelines, each explicitly enriched with graph features, with two state-of-the-art graph neural networks (GNNs). To ensure a rigorous comparison, Long Short-Term Memory (LSTM) and Attention-based Bidirectional LSTM networks are adopted as high-capacity baselines, as they have previously achieved top-ranked results. Experiments on ten diverse, real-world datasets reveal a consistent pattern: Even lightweight models, once augmented with carefully encoded graph information, outperform sophisticated deep architectures that neglect such relational cues. These findings underscore the decisive impact of graph encoding over model complexity in outcome-oriented prediction and provide actionable guidance for practitioners selecting predictive pipelines.