[Paper Review] (Blue) Taxi Destination and Trip Time Prediction from Partial Trajectories

Hoang Thanh Lam, Ernesto Diaz-Aviles+3 more|arXiv (Cornell University)|Sep 17, 2015

Human Mobility and Location-Based Analysis8 references16 citations

TL;DR

This paper presents a data-driven approach for predicting taxi destinations and trip times from partial GPS trajectories using trip matching and ensemble learning. By leveraging historical trip patterns and aggregating predictions from multiple models, the method achieved 3rd place in the ECML/PKDD 2015 challenge for trip time prediction and 7th for destination prediction, demonstrating high robustness and consistency across test sets.

ABSTRACT

Real-time estimation of destination and travel time for taxis is of great importance for existing electronic dispatch systems. We present an approach based on trip matching and ensemble learning, in which we leverage the patterns observed in a dataset of roughly 1.7 million taxi journeys to predict the corresponding final destination and travel time for ongoing taxi trips, as a solution for the ECML/PKDD Discovery Challenge 2015 competition. The results of our empirical evaluation show that our approach is effective and very robust, which led our team -- BlueTaxi -- to the 3rd and 7th position of the final rankings for the trip time and destination prediction tasks, respectively. Given the fact that the final rankings were computed using a very small test set (with only 320 trips) we believe that our approach is one of the most robust solutions for the challenge based on the consistency of our good results across the test sets.

Motivation & Objective

• To improve electronic taxi dispatching by predicting final destinations and trip durations in real time using incomplete GPS trajectories.
• To address the challenge of unreliable or missing destination reporting by taxi drivers in urban mobility systems.
• To develop a robust, data-driven solution for real-time prediction that performs consistently across diverse traffic and trajectory conditions.
• To outperform existing methods in the ECML/PKDD 2015 Discovery Challenge on both destination and trip time prediction tasks.
• To validate the effectiveness of trip matching and ensemble learning on a large-scale dataset of 1.7 million taxi trips in Porto, Portugal.

Proposed method

• The method uses trip matching to identify historical taxi trips with similar partial trajectories to the current test trip.
• Features are extracted from matched historical trips, including destination and trip time, which serve as training signals for predictive models.
• Random Forest and Extremely Randomized Trees models are trained on the matched features, with hyperparameters tuned for optimal performance.
• Multiple model ensembles are created using averaging and regularized linear regression (L1 and L2 penalties) to combine predictions and improve robustness.
• The training data is filtered to include only trips with 2 to 612 GPS updates, matching the distribution observed in the test set.
• Final predictions are generated using ensemble averaging or meta-regression, with model selection based on public leaderboard feedback.

Experimental results

Research questions

• RQ1 Can historical trip patterns with similar partial trajectories be reliably used to predict the final destination of an ongoing taxi trip?
• RQ2 How effective is ensemble learning in combining predictions from multiple models trained on matched historical trips for trip time and destination prediction?
• RQ3 To what extent does model robustness and consistency across test sets correlate with final competition ranking in real-world prediction challenges?
• RQ4 How do different ensemble strategies—averaging, L1, and L2 regularization—affect predictive performance on unseen test data?
• RQ5 Can a model trained on a subset of trips with limited GPS updates (2–612 points) generalize well to the full test set?

Key findings

• The model ensemble using averaging of four model types (ME5) achieved the best public leaderboard score of RMSLE = 0.49408, securing 3rd place in the trip time prediction task.
• The ensemble using L1-regularized linear regression (Lasso) achieved the best average performance across both public and private test sets with RMSLE = 0.51173.
• The model ensemble ME2, which used L2 regularization, achieved a private leaderboard RMSLE of 0.51327—ranking 1st if selected, but not chosen due to public leaderboard feedback.
• The method demonstrated high robustness, with consistent performance across both public and private test sets, indicating strong generalization.
• The approach achieved 7th place in the destination prediction task, confirming its effectiveness for both prediction objectives.
• Preprocessing steps effectively handled missing GPS points and erroneous timestamps, improving model reliability.