Intelligent Traffic Path Recommendations Using Machine Learning based Temporal Feature Aggregation and Time Series Based Graph Routing

Abstract

Urban transportation networks generate massive volumes of spatio-temporal data, and the effective exploitation of such data to enable intelligent route recommendations has become a critical priority in smart city development. Existing systems predominantly rely on static or short-horizon prediction models that fail to capture the complex, non-stationary dynamics characterizing real-world vehicular traffic, particularly under congestion-prone conditions in densely populated urban corridors. The challenges are further compounded by high-dimensional, noisy, and heterogeneous sensor data that makes robust feature extraction and temporal modeling particularly difficult, as traditional approaches exhibit poor generalization across varying prediction horizons and traffic regimes. This research proposes an integrated framework for traffic path recommendation that combines systematic feature extraction, temporal prediction, and dynamic graph-based routing. The methodology begins with the Feature Extraction for Vehicular Traffic Time Series (FEVTTS) algorithm, which applies Mutual Information, Random Forest Regression, and Correlation Coefficient analysis to derive an optimal, low-redundancy feature set from raw sensor data. The refined features are then fed to the Temporal Traffic Feature Aggregation Algorithm (TFTAA), which integrates short-term and long-term traffic patterns using a rolling window mechanism to produce accurate multi-horizon forecasts. Finally, the predicted traffic variables are embedded as dynamic edge weights into a Time Series Based Graph (TSBG), on which Dijkstra's algorithm computes optimal time-dependent routes. The proposed TFTAA algorithm employs a hybrid statistical and ensemble machine learning design, incorporating SARIMAX for seasonal decomposition and gradient-boosted residual correction using XGBoost, augmented by quantile regression for uncertainty quantification. The framework was validated using the London Traffic Dataset (Highways England) and a locally collected Pune city dataset comprising 26,785 and 25,920 samples, respectively, across diverse road types. TFTAA consistently outperformed SARIMAX and XGBoost baselines at all evaluated prediction horizons from 5 to 60 minutes, achieving a Mean Absolute Error (MAE) of 2.1 at the 5-minute horizon and 14.3 at the 60-minute horizon, compared to 25.1 for SARIMAX at the same horizon. The TSBG-based routing reduced average travel time by an average of 18.3% over static Dijkstra across 15 test days on the London dataset. The proposed framework demonstrates that tightly coupling predictive traffic modeling with dynamic graph routing yields measurable improvements in route efficiency over baseline approaches. The research advances the state of the art in intelligent transportation systems by providing a scalable, feature-engineered, and prediction-integrated path recommendation solution applicable to both developed and emerging urban road networks.