Deep Learning Approach for Route ETA Prediction in Navigation

Driving navigation is a core scenario of digital maps. When users request navigation, the map service provides three route options and displays the estimated time of arrival (ETA) for each. Accurate ETA prediction improves travel experience.

The article explores the use of deep learning models to predict segment‑level ETA and aggregate them into a full‑route ETA.

2. Exploration and Practice of Deep Learning Models

2.1 Model Selection

Traditional machine‑learning regressors such as Linear Regression, Random Forest (RF) and Gradient Boosting Decision Trees (GBDT) are commonly used for ETA prediction. While they are simple and effective, they have two main drawbacks: (1) model performance heavily depends on manually engineered features, and (2) they ignore upstream‑downstream road segment dependencies, leading to uncertainty.

Deep learning can address both issues. For example, historical speed information is aggregated over 10‑minute intervals for each weekday, but its selection depends on the estimated arrival time at upstream segments, introducing uncertainty.

To capture the sequential relationship of road segments, a Recurrent Neural Network (RNN) is adopted, and because of long‑term dependency issues, an LSTM architecture is chosen. The LSTM’s gate structure allows the model to learn which upstream features to retain for prediction.

2.2 Network Architecture

The overall framework consists of two parts: an LSTM module that predicts ETA for each road segment, and a stack of fully‑connected layers that aggregates segment ETAs and route‑level features to produce the final route ETA.

2.3 Segment‑Level ETA Prediction

For each segment t, the input vector X t contains real‑time traffic, historical traffic, and static segment features. The LSTM processes the sequence of segments, leveraging its temporal modeling capability.

2.4 Full‑Route ETA Prediction

After obtaining the cumulative ETA from the LSTM, static route attributes (e.g., length, departure time, peak‑hour flag) are concatenated and fed into fully‑connected layers to output the final route ETA.

The loss function is the standard mean squared error (MSE):

where N is the number of routes, ETA j is the predicted ETA for route j, and the ground‑truth is the actual travel time.

3. Model Effectiveness

Model performance is evaluated by ETA accuracy, emphasizing tolerance for overestimation (longer ETA) over underestimation. In experiments on Beijing data, the proposed model improved ETA accuracy and reduced MSE by 28.2%.

4. Conclusion

Introducing deep learning into navigation ETA prediction overcomes limitations of linear models and opens opportunities for incorporating richer features such as historical speed uncertainty, temporal patterns, incidents, and road‑network topology.