gps2topo-poi 0.1.3

A Python package to use GPS data of public transit routes and append topographical and Points of Interests (POI) related features.

Explainable Bus Arrival Time Prediction Model with Improved Features Related to Topography and Points of Interest

Usage

Step 1: Install the library

pip install gps2topo-poi

Step 2: Import the library

import gps2topo_poi as g2p

# dwell time POIs extraction
g2p.dwell_time_stop_classifier_poi_extraction.complete_extraction_pipelinecomplete_extraction_pipeline(bus_stops_path, segment_id_start=1)

# running time POIs extraction
g2p.running_time_poi_extraction.complete_extraction_pipeline(bus_stops_path, route_points_path, segment_id_start=1)

# Topographical feature extraction
g2p.topological_feature_extraction.complete_extraction_pipeline(bus_stops_path, route_points_path, segment_id_start=1)

Abstract

Accurate and reliable prediction of bus arrival times enhances passenger mobility experience. This study addresses a significant research gap by focusing on the complexities of predicting bus arrival times in heterogeneous traffic conditions. Unlike conventional prediction models, this research identifies hidden features related to topographical and Points of Interest (POIs) data, recognizing their critical role in reasoning. The methodology involves a two-fold approach, segmenting predictions into running time within a segment and dwell time at bus halts, using the multi-model ensemble technique. The results indicate that incorporating the new features (5 topographical and 10 POIs-related) has improved model performance by a reduction in MAE of 1.37 seconds (dwell time) and a decrease in MAPE by 0.7% (running time). While the enhancements in accuracy may appear modest, our focus lies on examining the influence of new features, offering valuable insights into the factors that cause delays. Moreover, we developed a dashboard showcasing real-time bus arrival times and highlighting delay reasoning using explainable AI techniques.

Results

Performance of Running Time Prediction with and without Topographical and POIs Data

XGBoost Model	RMSE (s)	MAE (s)	MAPE (%)	R2
Without topographical & POIs features	57.55	37.19	23.11	0.76
With topographical & POIs (total 25 features)	57.14	36.64	22.06	0.76
With topographical & POIs (total 19 features)	56.99	36.56	22.12	0.76

Performance of Dwell Time Prediction with and without POIs Data

XGBoost Model	RMSE (s)	MAE (s)	R2
Without POIs features	38.13	19.77	0.14
Original features with total_poi_count	38.2	19.6	0.14

Authors

• A.K. Warnakulasuriya - Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka
• C.D.R.M. Weerasinghe - Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka
• H.K.G.V.L. Wickramarathna - Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka
• Shiveswarran Ratneswaran - Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka
• Dr. Uthayasanker Thayasivam - Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka