The optimization of supervised machine learning models performancerepresents a constant challenge, especially in contexts with high-dimensionaldatasets or numerous correlated attributes. In this study, we propose a methodfor enriching tabular datasets, based on the use of statistics derived from agraph constructed from the similarity between instances in the dataset, aimingto capture structural correlations among the data. Instances take on the role ofvertices in the graph, while connections between them reflect their similarity.The original feature set (FO) is enriched with statistics extracted from thegraph (FG) to enhance the predictive power of machine learning models. Themethod was evaluated on ten public datasets from different domains, in twodistinct scenarios, across seven machine learning models, comparing predictionon the initial dataset (FO) with the dataset enriched with statistics extractedfrom its graph (FO+FG). The results revealed significant improvements inaccuracy metrics, with an average enhancement of approximately 4.9 percent. Inaddition to its flexibility for integration with existing enrichment techniques,the method presents itself as a effective alternative, particularly in situationswhere original datasets lack the necessary characteristics for traditional graph-based enrichment approaches.