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

Summary

The paper introduces RTGL, an open-source declarative domain-specific language for defining prediction tasks in Relational Deep Learning (RDL). RTGL is modeled on the proprietary Predictive Query Language (PQL) from Kumo.ai but is an independent implementation targeting open ecosystems such as RelBench and ReDeLEx. The pipeline parses RTGL queries with ANTLR4 into an AST, performs schema-aware semantic validation (including strict temporal checks separating past observation from future prediction windows), and compiles queries into temporally bounded SQL using CTEs. The authors reconstruct several RelBench tasks on rel-f1 and rel-stack and report that the comparison exposed genuine temporal-leakage defects in the official manual SQL: driver-dnf filtered drivers using future activity, and user-engagement/user-post-comment omitted registration-date filters, so baselines included users not yet registered. They also define new binary, multiclass, multilabel, and regression tasks on the enterprise-scale ctu-seznam (Seznam) dataset from ReDeLEx, and train GraphSAGE and HGT models on a selection of tasks, reporting e.g. test accuracy .7219±.0123 (SAGE) on driver-dnf and test AUPRC .6274±.0017 (SAGE) on client-services. Main limitations: downstream evaluation lacks external baselines and does not quantify how the corrected tables change benchmark scores; the number of random seeds behind the mean±std figures is unstated; multi-hop tasks required intermediate views; and no runtime or performance measurements are provided despite acknowledged scalability limits.