Shared temporal filters cut forecasting costs in time series data
STeP-TS: Shared Temporal Filters with Pruning for Time Series
Machine Learning
Summary
Predicting what will happen next in a set of related time series data is hard because different variables change in different ways over time. The authors introduce a simple model that uses a shared set of patterns to understand all variables, then adjusts how much each pattern matters for each specific variable. This approach makes the model smaller and faster but still quite accurate. They also developed a way to remove less important parts of the model to make it even leaner without losing much accuracy.
What this means in practice
- •For mobile app developers: Run fast and accurate time series predictions on-device with a lightweight model that adapts to multiple variables.$Commercial implications: Allows selling predictive features in resource-limited apps such as health monitors or financial tools requiring fast inference.
- •For industrial control engineers: Deploy compact forecasting models to monitor multiple sensor readings efficiently for timely maintenance or anomaly detection.
Authors
Hanne Raum, Jasper Hoffmann, Joschka Boedecker
Abstract
Multivariate time series forecasting requires capturing temporal dependencies across multiple scales while accounting for both shared and variable-specific dynamics. Compact forecasting models have emerged as an attractive alternative to increasingly large architectures, but preserving strong predictive performance within tight computational and memory budgets remains challenging. We propose a compact one-layer convolutional framework that learns a shared bank of multi-scale temporal filters of each variable. Lightweight variable-specific scalar weights determine the contribution of each temporal component, enabling the model to capture shared temporal structure while adapting to variable-specific dynamics. The learned weights and filter-output energy further provide a natural signal for structured pruning of convolutional filters, reducing computational complexity with minimal degradation in forecasting performance. We additionally introduce an amplitude-adaptive normalization mechanism that preserves variable-dependent temporal amplitudes. Experiments on multiple datasets demonstrate forecasting performance competitive with state-of-the-art methods while requiring substantially fewer parameters and inference FLOPs and achieving low per-sample latency.