Uber's DeepETT Boosts Traffic Forecasts

When you hail an Uber, the app instantly calculates the fastest route and estimates your arrival time. Both critical functions rely heavily on accurate, real-time traffic predictions. After a decade of relying on its existing system, Uber has overhauled its traffic forecasting stack with DeepETT (Deep Estimated Travel Time). This deep learning-powered system, detailed on the Uber Engineering blog, improves long-trip arrival time accuracy by 6% and boosts forecast variance explained by 19%. It's also a massive engineering feat, processing over 2 million forecasts per second.

Uber's traffic forecasting system transforms raw GPS data from millions of driver phones into predictions for road segment speeds over the next few hours. Billions of location updates daily are mapped onto a global road graph, generating precise measurements of segment traversal times. These forecasts are foundational, impacting routing, pricing, and driver navigation.

Designing for Scale and Accuracy

The previous system was stable and scalable but struggled with rapidly changing conditions, sparse data in less urban areas, and the compounding errors of long trips. DeepETT was built to address these limitations, requiring a system that could adapt quickly, generalize across diverse geographies, and leverage Uber's vast data volume. This made it a prime candidate for a deep learning rebuild.

Two key scoping decisions de-risked the project: first, DeepETT predicts segment-level travel times directly, decoupling it from the complex routing engine during training. Second, it uses fixed-size inputs derived from pre-aggregated data, avoiding the performance bottlenecks of dynamic graphs at scale. This approach maintains predictable inference while utilizing all available data.

Contracts and Performance Metrics

DeepETT operates under two defined contracts: segment-level forecasts every five minutes for up to three hours, measured by mean squared error (MSE), and trip-level arrival times that must exhibit lower error than the legacy system, measured by mean absolute error (MAE).

While segment-level MSE optimization is direct, trip-level MAE optimization is indirect, posing a challenge. Improvements in segment-level accuracy don't always guarantee improvements at the trip level due to compounding errors and calibration issues. This paradox led to a focus on both resolution (information content) and calibration (systematic error) in their metrics.

Model Architecture and Data Pipeline

The DeepETT model employs a transformer architecture, treating various spatiotemporal views of traffic data as tokens. These views include the segment itself, its neighborhood, and broader regions, combined with historical and real-time temporal data. Contextual features like time-of-week and holidays are also incorporated.

This design allows the model to dynamically weigh information sources, relying on recent segment data for busy roads and broader spatial aggregates for sparser areas. The transformer efficiently learns interactions between these views to produce forecasts conditioned on the requested horizon.

Robust production pipelines using Apache Spark and Apache Flink ensure real-time data ingestion and global serving. This infrastructure supports over 160,000 feature rows per second and the massive throughput of 2 million segment-level predictions per second.

The Calibration Challenge

A surprising challenge emerged post-launch: downstream arrival time models saw decreased accuracy despite DeepETT's improved forecasts. The root cause was calibration drift, where segment-level miscalibrations snowballed into larger trip-level errors, particularly varying by city and time of day.

This drift meant the model's forecasting contract wasn't consistently met, necessitating real-time calibration adjustments in downstream systems to fully realize DeepETT's benefits. This highlights the intricate interplay between granular predictions and end-user experience in complex systems like Uber traffic forecasting.

The successful deployment of DeepETT showcases Uber's prowess in scalable AI deployments, pushing the boundaries of machine learning in transportation.