Discovery of trend dependencies over time-series
Abstract
We improve constraint-based data quality using trend dependency (TD) discovery, extending existing order dependencies (ODs) to allow variations and exceptions. Unlike ODs, TDs capture approximate functional mappings between attributes, addressing the limitations of monotonicity. Our approach involves automatic discovery over entire datasets and piecewise subsets. Optimizing across all possible mappings is impractical, but a single linear pass enables efficient pruning, making segmentation and trend discovery feasible with minimal accuracy loss. We conducted comprehensive experiments on real-world and synthetic data to evaluate our models.