DSETA: A Dual-Stage Continual Learning Framework for Travel Time Prediction in Dynamic Traffic Environments
DSETA introduces a dual-stage continual learning framework for ETA prediction that separates inter-day (long-term) and intra-day (short-term) learning to handle dynamic traffic patterns The intra-day stage uses real-time data for rapid adaptation to sudden events like accidents and holidays, while the inter-day stage leverages aggregated historical windows to capture seasonal trends and network evolution A Historical Traffic Knowledge Consolidation module prevents catastrophic forgetting, preser
Analysis
TL;DR
- DSETA introduces a dual-stage continual learning framework for ETA prediction that separates inter-day (long-term) and intra-day (short-term) learning to handle dynamic traffic patterns
- The intra-day stage uses real-time data for rapid adaptation to sudden events like accidents and holidays, while the inter-day stage leverages aggregated historical windows to capture seasonal trends and network evolution
- A Historical Traffic Knowledge Consolidation module prevents catastrophic forgetting, preserving knowledge of regular traffic patterns during incremental updates
- Deployed in DiDi's production environment processing hundreds of millions of daily requests, with A/B tests showing MAE reductions of 6.62% (Beijing), 0.73% (Wuhan), and 2.40% (Xi'an)
- Addresses a critical gap in existing ETA systems that either fail to adapt to irregular patterns or cannot disentangle long-term trends from short-term fluctuations
Why It Matters
This work demonstrates how continual learning techniques can be practically applied to large-scale industrial prediction systems, bridging the gap between academic research and real-world deployment. For AI practitioners working on time-series forecasting or recommendation systems, DSETA's dual-stage approach offers a replicable architecture for handling both rapid distribution shifts and gradual long-term drifts simultaneously. The successful production deployment at DiDi validates the framework's scalability and robustness under extreme traffic volumes.
Technical Details
- Dual-Stage Architecture: The framework divides continual learning into two distinct stages — the intra-day stage processes real-time streaming data for immediate adaptation to short-term anomalies (accidents, weather events, holidays), while the inter-day stage operates on aggregated historical data within a sliding time window to model long-term distribution shifts such as seasonal patterns and infrastructure changes
- Historical Traffic Knowledge Consolidation Module: Designed to mitigate catastrophic forgetting by preserving representations of regular traffic patterns learned from historical data, ensuring that incremental updates do not erase previously acquired knowledge of stable traffic behaviors
- Incremental Update Mechanism: Rather than full retraining, DSETA employs an incrementally updated approach that continuously absorbs new data distributions while maintaining performance on established patterns, making it computationally efficient for high-throughput environments
- Industrial-Scale Validation: Evaluated through extensive offline experiments and online A/B tests across three major Chinese cities (Beijing, Wuhan, Xi'an) on DiDi's real-world ride-hailing platform data, demonstrating consistent MAE improvements across diverse urban traffic environments
- Deployment Scale: Successfully deployed in DiDi's production system handling hundreds of millions of daily ETA requests, confirming the framework's operational viability at massive scale
Industry Insight
- The dual-stage continual learning paradigm is directly transferable to other domains requiring simultaneous adaptation to short-term volatility and long-term drift, such as demand forecasting, anomaly detection in IoT networks, and financial time-series prediction
- The Historical Traffic Knowledge Consolidation approach offers a practical blueprint for addressing catastrophic forgetting in production ML systems, where full retraining is infeasible due to computational cost and data latency constraints
- The DiDi deployment case study underscores the importance of online A/B testing across geographically diverse markets before full rollout, as performance gains varied significantly (0.73% to 6.62% MAE reduction) depending on local traffic dynamics and data characteristics
Disclaimer: The above content is generated by AI and is for reference only.