ETA Engine: A New York Taxi Trip Time Prediction System Based on Neural Network and LightGBM Fusion

The open-source project eta-engine builds a high-precision trip time prediction system based on the New York taxi dataset by fusing a neural network and LightGBM integrated model, combining features such as regional data, timestamps, and passenger count. It solves the problem that traditional methods struggle to capture complex non-linear relationships, provides technical support for intelligent travel, and demonstrates a typical paradigm of integrating deep learning with traditional machine learning.

In modern urban transportation systems, taxi and ride-hailing services are important parts of daily travel. However, accurate trip time estimation is affected by various complex factors such as road congestion, weather, time periods, and regional attributes. Traditional rules or simple statistical methods struggle to capture non-linear relationships, leading to large deviations. The New York City taxi dataset provides real data for research, and the eta-engine project is committed to building a high-precision prediction engine that integrates deep learning and traditional ML technologies to support intelligent travel.

eta-engine adopts a hybrid model architecture:

1. Neural Network Module: Processes multi-dimensional structured inputs (geographic coordinates, regional codes, time features, passenger count, etc.), and automatically extracts high-order features and implicit correlations (such as time-region interactions, holiday impacts).
2. LightGBM Integration Module: Receives outputs from the neural network and original features, corrects residuals by combining weak learners through gradient boosting, and achieves efficient and accurate prediction.
3. Feature Engineering: Deeply mines cross-combinations of regional data (New York taxi operation area division) and time features to capture fine-grained spatio-temporal patterns; passenger count reflects the impact of vehicle load.

Training challenges and solutions:

• Data quality: Clean outliers, missing values, and incorrect records to ensure input quality.
• Class imbalance: There are many short-trip samples; weighted loss or sampling strategies are used to correct bias. Evaluation metrics: RMSE, MAE, R² scores; generalization ability is ensured through cross-validation and a held-out test set. Integration advantages: Neural networks provide strong representation learning, while LightGBM provides interpretable feature importance analysis.

Practical value:

• Ride-hailing platforms: Optimize dispatch algorithms, reduce passenger waiting time, and improve experience.
• Taxi drivers: Plan order-receiving strategies and choose optimal routes during peak hours.
• Urban planners: Reveal traffic bottlenecks and support infrastructure improvement. Technical significance: Demonstrates the fusion paradigm of deep learning and traditional ML, solving problems such as high training cost and poor interpretability of pure neural networks, and the inability of pure tree models to capture complex interactions, providing reference for similar scenarios.

Project highlights: Multi-source feature fusion, complementary advantages of deep learning and gradient boosting, and specialized modeling of spatio-temporal data. Future directions: Introduce real-time traffic flow to improve timeliness, use attention mechanisms to enhance key feature capture, and migrate to other cities/scenarios. Learning value: Provides ML engineering practitioners with a complete process reference from data preprocessing to model integration.