Machine learning (ML) models frequently struggle in domains where labeled data is limited or sensitive. To address this challenge, we explore a semantic-driven data augmentation approach that incorporates external knowledge into tabular datasets. Our method leverages neuro-symbolic techniques to enrich training data with structured context derived from knowledge graphs (KGs), aiming to enhance the predictive capabilities of standard ML algorithms. We evaluate multiple approaches for integrating KG information into ML pipelines and examine their impact on model performance across binary classification tasks involving medical datasets such as heart disease and chronic kidney disease. The experimental setup includes four ML models and four distinct KG embedding algorithms, with performance evaluated using accuracy and F2 score. Results show that augmenting tabular features with semantic distance metrics from KG embeddings yields notable improvements. For instance, XGBoost achieves a significant F2 score increase from 75.19% to 90.85% in heart disease prediction. These results suggest that semantic augmentation of tabular datasets has the potential to enhance ML prediction1.
Methodology To enrich tabular data with KG information, we propose a pipeline that integrates KG embeddings into ML models (see Figure 1). The process begins by defining domain ontologies to represent tabular dataset’s features and constructing KGs using mappings from tabular features to ontology concepts and relations. These KGs are then embedded into vector spaces using diferent KG embedding methods. We explore several methods for computing information in the embedding space which is then used to augment tabular data with KG-derived features to enhance ML performance, as illustrated in Figure 2. These strategies difer in how they leverage the semantic information encoded in KG embeddings. One method, EmbedOnly, uses only the embeddings to assess whether the learned semantic structure can replace raw features. In EmbedAugTab, embeddings are combined with the original tabular features to enrich the data with latent relational information. The DistAugTab approach introduces distance-based features: for each instance, we calculate the distance from its embedding to class centroids, capturing proximity-based semantics. This is extended in EmbedDistTabAug, which integrates both the raw embeddings and their distance-based metrics into the dataset. To capture higher-order semantics, ClusterAugTab applies clustering in the embedding space, assigning each instance a cluster membership to reflect latent groupings. EmbedClusterAugTab builds on this by combining cluster memberships with the full embeddings for even more expressive augmentation. Finally, we investigate feature interaction strategies. In InteraAugTab and EmbedInteraAugTab, we compute element-wise interactions between original features and embeddings, enabling the model to learn complex joint efects between raw clinical measurements and semantic knowledge from the KG.
Our experiments apply these strategies to binary classification tasks in the healthcare domain (heart disease and chronic kidney disease), using four ML models (KNN, SVM, XGBoost and a feedforward neural network (NN)). Performance is evaluated using accuracy and F2 score.
Results Our evaluation covered eight KG-based data augmentation strategies applied to binary classification tasks in heart disease and chronic kidney disease prediction. In Table 1, 1 we present the average accuracy and F2 scores across diferent embedding methods, classifiers, and augmentation strategies for heart disease prediction. Similar trends were observed in the kidney disease experiments, where integrating KG-based information also led to improvements in predictive performance. Among the evaluated strategies, those that incorporated distance-based features from the embedding space (e.g., DistAugTab) outperformed other approaches.
Among the embedding methods, RDF2Vec emerged as the most efective overall, yielding
stable and strong results across both tasks and models, likely due to its ability to capture semantic
paths in the KG. Node2Vec also performed well in scenarios where local graph structure was more
informative, e.g., improving KNN in kidney disease and XGBoost in heart disease prediction.
Notably, XGBoost, despite achieving the highest F2 score in some configurations, showed
1Due to space limitations, we report results only for heart disease prediction in this extended abstract. Full results,
including those for chronic kidney disease and additional scenarios, are available in [
These findings support the hypothesis that semantic enrichment, particularly through embedding-derived features, can enhance ML performance on tabular data, especially in domains with limited or sensitive datasets.
Conclusion This work shows that semantically enriching tabular data using KG embeddings can enhance the predictive performance of ML models, particularly in low-data medical scenarios. Integrating structured domain knowledge through augmentation strategies, especially distancebased features, led to notable F2 score gains. While promising for context-aware predictions in sensitive domains, challenges remain in generalizability. Future work will explore the efectiveness of KGs across diverse, data-scarce domains to address ML models’ data dependency. Acknowledgments This work was supported by the FFG SENSE (894802) and FAIR-AI (904624) projects, as well as by the Austrian Science Fund (FWF) BILAI 10.55776/COE12 and HOnEst (V 745-N) projects. For open access purposes, the author has applied a CC BY public copyright license to any author accepted manuscript version arising from this submission. Declaration on Generative AI During the preparation of this work, the author(s) used GPT-4o in order to brainstorm ideas about the title. After using these tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.