🤖 machine learning

DSETA: A Dual-Stage Continual Learning Framework for Travel Time Prediction in Dynamic Traffic Environments

The paper proposes DSETA, a dual-stage continual learning framework that integrates intra-day real-time adaptation and inter-day historical consolidation to effectively address dynamic traffic patterns and prevent catastrophic forgetting, achieving significant accuracy improvements in DiDi's production ETA prediction system across multiple major cities.

Original authors: Yanming Lyu, Yue Cheng, Lingkun Li, Ruipeng Gao, Xinyue Liu, Hui Gao, Qiang Ni

Published 2026-08-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yanming Lyu, Yue Cheng, Lingkun Li, Ruipeng Gao, Xinyue Liu, Hui Gao, Qiang Ni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess how long it will take to get from your house to a friend's party. In a perfect world, traffic would be like a metronome: steady, predictable, and the same every day. But in the real world, traffic is more like a chaotic dance party. Sometimes a surprise accident blocks the road, sometimes a holiday brings a flood of extra cars, and sometimes a new road opens up, changing the whole rhythm of the city. This is the world of ETA (Estimated Time of Arrival) prediction, a super-important job for ride-hailing apps like DiDi or Uber. If the app guesses wrong, you might be late, or the driver might get stuck in a traffic jam they didn't expect.

For a long time, computers tried to solve this by looking at history. They'd say, "Last Tuesday at 5 PM, it took 20 minutes, so it will take 20 minutes today." But this approach is like trying to navigate a stormy ocean with a map from last year; it works fine when the weather is calm, but it fails miserably when a sudden storm hits. This is where Continual Learning comes in. Think of this as a brain that doesn't just memorize facts but keeps learning new things every day without forgetting the old ones. The big challenge is balancing two things: remembering the usual patterns (like rush hour every morning) while also instantly reacting to weird, one-time events (like a parade or a crash). If the computer learns too much about the parade, it might forget that rush hour usually happens at 5 PM. If it learns too much about the usual routine, it won't know the parade is happening right now.

This is exactly the puzzle the researchers at Beijing Jiaotong University and DiDi set out to solve with a new framework called DSETA. They realized that traffic has two different "speeds" of change: fast changes that happen within a single day, and slow changes that happen over weeks or months. To handle this, they built a two-stage learning system that acts like a driver with two different sets of eyes.

First, there's the Intra-day stage (the "Day-Of" eye). This part of the system is hyper-alert to what's happening right now. It looks at real-time data, minute by minute, to catch sudden surprises. If a car crashes or a holiday traffic surge starts, this stage instantly tweaks the prediction. It's like a driver who sees a red light ahead and immediately slows down, even if the map said the road was clear five minutes ago. To make sure this quick reaction doesn't make the driver forget how to drive normally, they use a clever trick called Parameter Isolation. Imagine the driver's brain has two parts: a "Steady Hand" that knows the basic rules of the road (like how long a specific street usually takes) and a "Quick Reflex" that handles sudden swerves. The system freezes the "Steady Hand" so it doesn't get confused by the chaos, while only letting the "Quick Reflex" learn from the new, messy data.

Second, there's the Inter-day stage (the "Long-Term" eye). This part wakes up once a day to look at the bigger picture. It asks, "How has the city changed over the last few weeks?" Maybe a new road opened, or the seasons are changing, shifting when people usually go to work. This stage uses a method called Weekly Information Replay, which is like reviewing a highlight reel of the past few weeks. It makes sure the system remembers that Tuesdays are usually busy, even if today was a weird, quiet Tuesday. It also uses Knowledge Distillation, a fancy way of saying it checks its new guesses against its old, reliable knowledge to make sure it doesn't accidentally unlearn the basics.

The researchers tested this system on real data from DiDi, covering millions of trips in cities like Beijing, Wuhan, and Xi'an. They didn't just run simulations; they actually put DSETA into the real app and let it predict trips for real users. The results were impressive. In Beijing, the system reduced the average prediction error by 6.62%. In Wuhan, it cut the error by 0.73%, and in Xi'an, by 2.40%. More importantly, it made the app much more reliable during the chaotic times. The "bad case" rate—where the prediction was wildly wrong—dropped significantly, meaning fewer drivers getting stuck in unexpected jams and fewer passengers waiting longer than expected.

What's really cool is that this system is robust. Even if the computer gets busy and misses a few quick updates during the day, the system doesn't crash or give terrible advice. It has a safety net that keeps using the last good prediction until the new one is ready. The researchers found that this dual-stage approach works better than older methods that tried to learn everything at once. By separating the "fast" learning from the "slow" learning, DSETA manages to be both flexible and stable.

In the end, DSETA shows that the best way to predict the future in a chaotic world isn't to try to memorize every single detail, but to have a system that knows when to pay attention to the now and when to remember the past. It's a bit like a seasoned taxi driver who knows the city's usual rhythm but also knows how to react instantly when a parade blocks the street. Thanks to this framework, ride-hailing apps can now give you a time estimate that is not just a guess based on history, but a smart, living prediction that adapts to the city as it breathes and changes.

Technical Summary: DSETA

Problem Statement

Estimated Time of Arrival (ETA) prediction is a cornerstone of intelligent transportation systems, critical for ride-hailing dispatch, route planning, and dynamic pricing. However, maintaining high accuracy in large, dynamic cities presents significant challenges due to the non-stationary nature of traffic. Existing deep learning models, while effective at capturing spatiotemporal dependencies, often remain static during inference, leading to suboptimal performance when faced with distribution shifts.

The authors identify three specific industrial challenges that current Continual Learning (CL) methods fail to address adequately:

Uncertain Dynamic Traffic Patterns: Current models struggle to predict ETAs under irregular conditions (e.g., accidents, holidays) which cause rapid, significant deviations from historical norms. On the DiDi platform, such irregular congestion accounted for approximately 60% of major prediction failures in 2024.
Mixed Multi-scale Temporal Patterns: Existing frameworks often fail to distinguish between short-term fluctuations (transient events like accidents) and long-term distribution shifts (seasonal trends, road network evolution). Treating these heterogeneous signals uniformly obscures true traffic dynamics.
Catastrophic Forgetting: Standard update strategies risk either failing to learn new distributions or overfitting to recent data, causing the model to forget foundational, recurring traffic patterns essential for daily operations.

Methodology: DSETA Framework

To address these challenges, the authors propose DSETA (Dual-Stage ETA), an incrementally updated framework that decomposes continual learning into two complementary stages: Intra-day and Inter-day updates.

1. Base Model Architecture

The underlying predictor is based on a DeepFM architecture, chosen for its effectiveness in modeling spatiotemporal correlations. It consists of:

Wide Module: Learns global and route-level features (e.g., driver ID, day-of-week, OD pair, weather) via embedding projections and generalized linear layers.
Deep Module: Models fine-grained, link-level information (e.g., link ID, length, congestion status) using a two-layer fully connected network.
Output: A 3-layer MLP combines representations from both modules to predict travel time.

2. Dual-Stage Continual Learning

The core innovation lies in separating the learning process by temporal scale:

Intra-day Stage (Short-term Adaptation):
- Goal: Rapidly adapt to short-term traffic fluctuations caused by transient events (accidents, holidays).
- Mechanism: Uses real-time streaming data collected in hourly windows within a single day.
- Parameter Isolation Strategy: To prevent catastrophic forgetting of stable patterns, the model parameters are split into a frozen component ( $u$ , the Wide module) and an adaptive component ( $v$ , the Deep module). Only the Deep module is updated using the current data batch, preserving the stability of global features while allowing sensitivity to local, real-time dynamics.
Inter-day Stage (Long-term Adaptation):
- Goal: Capture long-term distribution shifts such as seasonal trends and road network evolution.
- Mechanism: Occurs at the transition between days, utilizing aggregated historical data from a short lookback window (e.g., 3-5 weeks).
- Weekly Information Replay: To mitigate forgetting, the training dataset is constructed by replaying data from the same weekday (target subset) alongside data from other days (generic subset). This reinforces persistent temporal structures.
- Knowledge Distillation: The loss function includes a distillation term that penalizes divergence from the previous day's model state, ensuring smooth adaptation to long-term shifts without losing historical knowledge.

3. System Architecture & Fault Tolerance

The system is designed for high-concurrency industrial deployment (e.g., QPS $\ge$ 10k). It employs a Service Fault-Tolerance Mechanism featuring:

Parameter Buffering: Intermediate parameters are stored in a temporary buffer, not affecting live service.
Atomic Updates: Model switching occurs only if the parameter deviation is within a safe threshold; otherwise, the system reverts to the previous stable model.
Timeout Recovery: If training exceeds a time limit, the system automatically rolls back to the last known good state.

Key Contributions

DSETA Framework: A novel continual learning framework that explicitly captures both regular and irregular traffic patterns by decoupling learning into inter-day and intra-day stages.
Dual-Stage Update Mechanism: A specific design that distinguishes long-term distribution shifts (handled via inter-day updates with knowledge distillation) from short-term variations (handled via intra-day updates with real-time data).
Knowledge Consolidation Module: A combination of Weekly Information Replay (to preserve weekly structural trends) and Parameter Isolation (to protect stable global features during volatile updates) to effectively mitigate catastrophic forgetting.
Industrial Validation: Extensive offline and online experiments on real-world datasets from DiDi's platform, demonstrating robustness across multiple cities and traffic conditions.

Experimental Results

The authors evaluated DSETA using offline datasets from Beijing and online A/B tests across Beijing, Wuhan, and Xi'an.

Offline Performance (Beijing): DSETA outperformed state-of-the-art baselines (including WDR, HierETA, and iETA) across all metrics.
- MAE: Reduced to 110.21s (vs. 113.14s for iETA).
- Bad Case Rate: Reduced to 3.72% (vs. 4.07% for iETA), indicating significantly improved reliability under challenging conditions.
- 2-Min Accuracy: Achieved 69.70%.
- Ablation Studies: Confirmed that removing either the intra-day or inter-day stage, or the knowledge consolidation mechanisms, led to measurable performance degradation, validating the necessity of the dual-stage design.
Online A/B Tests: Conducted over two weeks (including a 6-day holiday period) across three cities.
- Beijing: MAE reduction of 6.62%.
- Wuhan: MAE reduction of 0.73%.
- Xi'an: MAE reduction of 2.40%.
- User Experience: DSETA consistently reduced user complaint rates, with the most significant improvements observed during holidays when traffic patterns are most irregular.
- Efficiency: The framework maintained latency comparable to the baseline control group even under high query loads (up to 6000 QPS), proving the efficacy of the fault-tolerance mechanisms.

Significance

The paper claims that DSETA successfully bridges the gap between academic continual learning research and industrial deployment requirements. By explicitly separating short-term volatility from long-term structural shifts, the framework addresses the critical limitation of existing models that either fail to adapt to irregular events or suffer from catastrophic forgetting of regular patterns.

The authors emphasize that DSETA has been successfully deployed in DiDi's production environment, processing hundreds of millions of daily requests. This deployment validates the framework's ability to maintain high prediction accuracy and system stability in a real-world, high-throughput setting, directly enhancing user experience and platform efficiency in dynamic traffic environments.