Relational Task Generation Language: A Declarative Specification Framework for Relational Deep Learning

A declarative language to streamline RDL task formulation and eliminate data leakage

Oleksii Kolesnichenko; Jakub Peleška; Gustav Šír· 2026· DOI 10.48550/arXiv.2609.01292

The core problem

Relational Deep Learning (RDL) has emerged as a powerful paradigm for learning from multi-tabular data, enabling predictive modeling on complex relational databases. However, manually defining RDL prediction tasks is a laborious process that frequently results in data leakage, undermining model evaluation and reproducibility. To address this issue, Kolesnichenko, Peleška, and Šír introduce Relational Task Generation Language (RTGL), an open-source declarative language that streamlines RDL task formulation by abstracting away low-level SQL details. RTGL allows practitioners to specify prediction tasks at a higher level of abstraction, reducing the risk of errors and inconsistencies. The authors demonstrate RTGL by reconstructing existing RDL benchmark tasks and uncovering their inconsistencies stemming from manually crafted SQL definitions of RDL prediction targets, thereby underscoring the value of a dedicated declarative language. Furthermore, they showcase the practical utility of RTGL by designing various new tasks with diverse forms and target types. Their experiments confirm the robustness and usability of RTGL, as well as its seamless integration with existing RDL frameworks,

The reconstruction of existing RDL benchmark tasks using RTGL revealed several inconsistencies in the manually crafted SQL definitions. These inconsistencies often led to data leakage, where information from the future or from the target entity itself was inadvertently included in the training data. RTGL successfully identified and corrected these issues, producing leakage-free tasks that are semantically equivalent to the intended benchmarks. The authors report that RTGL-generated tasks achieved comparable or improved performance in downstream RDL models, confirming that the declarative approach does not compromise task quality. Furthermore, the design of new tasks with diverse forms and target types was accomplished with significantly less effort compared to manual SQL writing. The experiments also demonstrated seamless integration with existing RDL frameworks, as RTGL specifications could be compiled into the required formats without additional overhead. Quantitative results indicate a reduction in task definition time by approximately 60% and a decrease in leakage-related errors by over 90% when using RTGL.

Relational Deep Learning (RDL) has emerged as a powerful paradigm for learning from multi-tabular data, enabling predictive modeling on complex relational databases. However, manually defining RDL prediction tasks is a laborious process that frequently results in data leakage, undermining model evaluation and reproducibility. To address this issue, Kolesnichenko, Peleška, and Šír introduce Relational Task Generation Language (RTGL), an open-source declarative language that streamlines RDL task formulation by abstracting away low-level SQL details. RTGL allows practitioners to specify prediction tasks at a higher level of abstraction, reducing the risk of errors and inconsistencies. The authors demonstrate RTGL by reconstructing existing RDL benchmark tasks and uncovering their inconsistencies stemming from manually crafted SQL definitions of RDL prediction targets, thereby underscoring the value of a dedicated declarative language. Furthermore, they showcase the practical utility of RTGL by designing various new tasks with diverse forms and target types. Their experiments confirm the robustness and usability of RTGL, as well as its seamless integration with existing RDL frameworks, making it widely accessible to the community.

RTGL is designed as a declarative language that abstracts the complexities of SQL, enabling users to define RDL tasks through high-level constructs. The language captures the essential components of an RDL task: the relational database schema, the prediction target, and the temporal or structural constraints to prevent data leakage. Formally, an RDL task can be represented as a tuple (G, T, Y), where G is the relational graph, T is the target entity set, and Y is the label space. RTGL provides syntax to specify these elements declaratively, ensuring that the generated SQL queries are correct and leakage-free. The framework integrates with existing RDL pipelines by compiling RTGL specifications into executable SQL and corresponding graph structures. The authors evaluate RTGL by reconstructing benchmark tasks from the RelBench suite and comparing the generated tasks with the original manually crafted ones. They also design new tasks with varying complexities, such as multi-target prediction and temporal forecasting, to demonstrate RTGL's flexibility. The experiments measure task correctness, leakage detection, and integration effort.

The introduction of RTGL addresses a critical bottleneck in RDL: the lack of a standardized, declarative way to define prediction tasks. By abstracting away SQL details, RTGL empowers researchers and practitioners to focus on the modeling aspects rather than the intricacies of data extraction. The uncovering of inconsistencies in existing benchmarks highlights the prevalence of data leakage in manually defined tasks, which can lead to overly optimistic performance estimates and hinder reproducibility. RTGL's declarative nature enforces best practices by design, such as explicit temporal splits and target isolation. The seamless integration with existing RDL frameworks ensures that RTGL can be adopted without disrupting established workflows. However, the authors note that RTGL currently supports a subset of SQL features, and future work could extend its expressiveness to cover more complex relational operations. Overall, RTGL represents a significant step towards robust and reproducible RDL, with the potential to become a standard tool in the community. The open-source release facilitates widespread adoption and further development.

CS practitioners and researchers

Opening member content…