The emergence of heterogeneous data sources, high-velocity data, and changing inconsistency patterns, including missing values, duplication, schema conflicts, and delayed updates, has made the provision of data consistency in AI-driven Extract, Transform, Load (ETL) systems more difficult. Conventional rule-based validation frameworks are not powerful enough to encode complex relationships and time dependencies that are present in modern data pipelines. To overcome these shortcomings, a Temporal Graph-Based Embedding Model (TGBEM) is proposed in this paper that combines Graph Neural Networks (GNNs) in learning structural relationships with Temporal Convolutional Networks (TCNs) in learning time-varying variations. The suggested framework converts ETL data into graphical features to show the dependence between data entities and processes time series to detect changing inconsistencies over time steps. The model is tested on a mix of real-world benchmark datasets and a synthetically created ETL dataset that renders realistic conditions on the pipeline, such as multi-source conflicts and temporal drift. According to the results of the experiments, it is clear that the proposed TGBEM greatly surpasses the traditional and existing versions of deep learning, reaching an accuracy of 95.80%, precision of 94.95%, recall of 93.70%, F1-score of 94.32%, and AUC of 96.10%. Moreover, the model also exhibits lower detection latency, which renders this model applicable in near-real-time validation of ETL. The comparative and ablation studies also support the fact that the combination of graph-based and temporal learning components is vital in improving the detection performance. These results indicate that hybrid graph-temporal models can be successfully used to face the challenging problem of inconsistency in next-generation AI-guided ETL systems.
Ganapathy et al. (Fri,) studied this question.