Related Experiment Videos
Temporal-Aware Bidirectional Local Label Propagation for Semi-Supervised Time Series Forecasting Under Sparse
Summary
This study introduces bidirectional local label propagation (Bi-LLP), a novel semi-supervised learning method for time-series forecasting. Bi-LLP effectively leverages sparse labels to improve forecasting accuracy, outperforming existing methods.
Area of Science:
- Machine Learning
- Time Series Analysis
- Data Science
Background:
- Time-series data collection is rapidly expanding, but labeled data remains scarce due to high costs and complexity.
- Existing semi-supervised learning (SSL) methods are underexplored for time-series forecasting due to challenges in aligning sparse labels with temporal dependencies.
Purpose of the Study:
- To develop a scalable SSL method for time-series forecasting that addresses extreme label sparsity.
- To improve the utilization of unlabeled data by effectively propagating limited labels.
Main Methods:
- Introduced bidirectional local label propagation (Bi-LLP), a novel SSL approach for time-series forecasting.
- Employed a dual-branch transformer to model forward and backward temporal dependencies.
- Incorporated a temporal proximity metric (TPM) and temporal-refined attention for enhanced pattern capture.
Main Results:
- Bi-LLP demonstrated superior performance compared to state-of-the-art SSL methods in time-series forecasting.
- The method showed particular effectiveness under conditions of extreme label sparsity.
- Experiments on diverse benchmarks and a real-world case validated the approach's efficacy.
Conclusions:
- Bi-LLP offers a scalable and effective solution for time-series forecasting with sparse labels.
- The proposed method successfully preserves temporal structures and enhances unlabeled data exploitation.
- This work advances SSL applications in time-series forecasting, especially in data-scarce scenarios.
