Usman Ahmed
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Usman Ahmed has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Matthew J. Roorda
Department of Civil and Mineral Engineering, University of Toronto, Toronto, Ontario, Canada
5 shared publications - Junshi Xu
Department of Civil and Mineral Engineering, University of Toronto, Toronto, Ontario, Canada
2 shared publications - Rolf Moeckel
Department of Mobility Systems Engineering, Technical University of Munich, Munich, Germany
2 shared publications - Ana Tsui Moreno
Technical University of Munich, Munich, Germany
1 shared publication - Glareh Amirjamshidi
Department of Civil and Mineral Engineering, University of Toronto, Toronto, Ontario, Canada
1 shared publication - Jason Hawkins
Department of Civil Engineering, University of Calgary, Calgary, AB, Canada
1 shared publication
Latest Publications
- 3 July 2026
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
Residential Price Modeling Using Spatially Validated Machine Learning Methods: A Comparison Across Geographical Contexts
Land use (i.e., buildings) and transportation infrastructure are tightly coupled systems, such that residential property prices play a critical role in transportation planning. In a similar manner, transportation infrastructure influences property prices, such that accurate forecasts of both systems are related research problems. The objective of this study is to examine these interactions and evaluate the effectiveness of machine learning methods in modeling residential real estate prices across different urban contexts. Specifically, we first examine the impact of land and transportation infrastructure on residential real estate prices using Extreme Gradient Boosting (XGBoost) and Random Forest (RF) machine learning methods, and compare results between two cities representing diverse geographic and socioeconomic contexts. Second, we investigate the application of machine learning methods on spatial data and provide a comparison of non-spatial and spatial cross-validation on the performance of machine learning methods. We use SHapley Additive exPlanations (SHAP) values to study the impact of land use and transportation infrastructure on real estate prices. The models are applied, and results are compared between the Rawalpindi and Islamabad Metropolitan Area in Pakistan and the City of Toronto in Canada. We find that, despite differences in demographics and economic development, the two cities exhibit similarities in the effect of transportation infrastructure and local amenities on dwelling prices. Proximity to the major central city cores (i.e., downtowns) increases sale price. The effect of transportation infrastructure is differentiated, with high quality transit (e.g., subway and BRT) increasing and conventional bus stop proximity decreasing sale price, respectively. We confirm the previous finding in other fields that non-spatial cross-validation over-estimates the prediction accuracy of machine learning algorithms on spatially referenced datasets. We find that the XGBoost model has slightly higher performance than the RF model. We recommend careful use of machine learning methods in the case of spatial data specifically in modeling of land prices.
- 3 July 2026
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
Residential Price Modeling Using Spatially Validated Machine Learning Methods: A Comparison Across Geographical Contexts
Land use (i.e., buildings) and transportation infrastructure are tightly coupled systems, such that residential property prices play a critical role in transportation planning. In a similar manner, transportation infrastructure influences property prices, such that accurate forecasts of both systems are related research problems. The objective of this study is to examine these interactions and evaluate the effectiveness of machine learning methods in modeling residential real estate prices across different urban contexts. Specifically, we first examine the impact of land and transportation infrastructure on residential real estate prices using Extreme Gradient Boosting (XGBoost) and Random Forest (RF) machine learning methods, and compare results between two cities representing diverse geographic and socioeconomic contexts. Second, we investigate the application of machine learning methods on spatial data and provide a comparison of non-spatial and spatial cross-validation on the performance of machine learning methods. We use SHapley Additive exPlanations (SHAP) values to study the impact of land use and transportation infrastructure on real estate prices. The models are applied, and results are compared between the Rawalpindi and Islamabad Metropolitan Area in Pakistan and the City of Toronto in Canada. We find that, despite differences in demographics and economic development, the two cities exhibit similarities in the effect of transportation infrastructure and local amenities on dwelling prices. Proximity to the major central city cores (i.e., downtowns) increases sale price. The effect of transportation infrastructure is differentiated, with high quality transit (e.g., subway and BRT) increasing and conventional bus stop proximity decreasing sale price, respectively. We confirm the previous finding in other fields that non-spatial cross-validation over-estimates the prediction accuracy of machine learning algorithms on spatially referenced datasets. We find that the XGBoost model has slightly higher performance than the RF model. We recommend careful use of machine learning methods in the case of spatial data specifically in modeling of land prices.
- Open Access
- 15 May 2026
- Published by Elsevier in SSRN Electronic Journal
Incident clearance time; uncertainty quantification; bi-level forecasting; calibrated routing; spatial risk; Tennessee highways
Incident clearance prediction is useful only when it is both accurate and decision ready. Existing bi-level incident-duration models typically rely on hard routing between short and long regimes, producing point forecasts that are brittle near the classification boundary and provide little information about model confidence. This paper develops an uncertainty-aware bi-level forecasting framework using 723,614 Tennessee Statewide Computer-Aided Dispatch System (SWCS) highway incidents from 2020–2025, scaling the bi-level architecture to a six-year statewide corpus.On a stratified 20% hold-out, the Stage 1 classifier achieves weighted F1 = 0.733, outper- forming logistic regression (F1 = 0.618), confirming that nonlinear interactions are important for prediction performance. The regime-specific regressors show distinct performance: Stage 2a reaches R2log = 0.201 with MAE 8.42 min for short incidents, while Stage 2b reaches R2log = 0.455 and MAE 148.52 min for long incidents. The review-band policy achieves empirical 90% interval coverage of 0.911, flags 16.9% of incidents as uncertain, lowers false escalation to 0.67%, and reduces overall MAE from 97.74 to 97.12 min compared with hard routing. A spatial ablation isolates the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) feature contribution: a +0.022 improvement in R2log for the long regime with negligible short-regime effect, confirming that spatial context is regime-specific and that this result is robust across dataset scales. On the six-year corpus, DBSCAN achieves 98.2% incident coverage compared to 49.1% in the single-year study, confirming that denser parameterization is necessary for state-scale corpora.
- 15 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessIncident clearance time; uncertainty quantification; bi-level forecasting; calibrated routing; spatial risk; Tennessee highways
Incident clearance prediction is useful only when it is both accurate and decision ready. Existing bi-level incident-duration models typically rely on hard routing between short and long regimes, producing point forecasts that are brittle near the classification boundary and provide little information about model confidence. This paper develops an uncertainty-aware bi-level forecasting framework using 723,614 Tennessee Statewide Computer-Aided Dispatch System (SWCS) highway incidents from 2020–2025, scaling the bi-level architecture to a six-year statewide corpus.On a stratified 20% hold-out, the Stage 1 classifier achieves weighted F1 = 0.733, outper- forming logistic regression (F1 = 0.618), confirming that nonlinear interactions are important for prediction performance. The regime-specific regressors show distinct performance: Stage 2a reaches R2log = 0.201 with MAE 8.42 min for short incidents, while Stage 2b reaches R2log = 0.455 and MAE 148.52 min for long incidents. The review-band policy achieves empirical 90% interval coverage of 0.911, flags 16.9% of incidents as uncertain, lowers false escalation to 0.67%, and reduces overall MAE from 97.74 to 97.12 min compared with hard routing. A spatial ablation isolates the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) feature contribution: a +0.022 improvement in R2log for the long regime with negligible short-regime effect, confirming that spatial context is regime-specific and that this result is robust across dataset scales. On the six-year corpus, DBSCAN achieves 98.2% incident coverage compared to 49.1% in the single-year study, confirming that denser parameterization is necessary for state-scale corpora.
- 26 April 2025
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
Urban Air Pollution Data Collection, Mapping, and Prediction Using Mobile Sensors Installed on Courier Trucks
Particulate matter with a diameter of 2.5 microns or less (PM 2.5 ) has significant impacts on human health, making it essential to understand its spatial and temporal variations. This study focuses on developing land use regression (LUR) models and improving their performance in predicting PM 2.5 concentrations in an urban setting. In this study, air quality data were collected using a sensor on a courier truck in downtown Toronto. Extreme gradient boosting (XGBoost), a machine learning algorithm, was employed to address limitations in traditional linear regression based LUR models, incorporating predictors such as land use, meteorology, and emissions to build robust models. A total of 27 models were trained, with varying road segment lengths, predictors, and outlier treatment thresholds. Three models tested the impact of road segment length on model predictions. Eight models examined the effect of removing outliers with different thresholds, revealing that appropriate thresholds improve accuracy. Ten models assessed the addition of emission and traffic data, which did not enhance performance, likely because of overlapping effects with other predictors. In six models, time-variant predictors such as time of day, month, humidity, wind speed, temperature, and pollutant concentrations from stationary stations were included. Adding these predictors significantly improved model performance, highlighting the complex relationships in LUR models for PM 2.5 predictions and offering valuable insights for air quality assessment.
- 26 April 2025
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
Urban Air Pollution Data Collection, Mapping, and Prediction Using Mobile Sensors Installed on Courier Trucks
Particulate matter with a diameter of 2.5 microns or less (PM 2.5 ) has significant impacts on human health, making it essential to understand its spatial and temporal variations. This study focuses on developing land use regression (LUR) models and improving their performance in predicting PM 2.5 concentrations in an urban setting. In this study, air quality data were collected using a sensor on a courier truck in downtown Toronto. Extreme gradient boosting (XGBoost), a machine learning algorithm, was employed to address limitations in traditional linear regression based LUR models, incorporating predictors such as land use, meteorology, and emissions to build robust models. A total of 27 models were trained, with varying road segment lengths, predictors, and outlier treatment thresholds. Three models tested the impact of road segment length on model predictions. Eight models examined the effect of removing outliers with different thresholds, revealing that appropriate thresholds improve accuracy. Ten models assessed the addition of emission and traffic data, which did not enhance performance, likely because of overlapping effects with other predictors. In six models, time-variant predictors such as time of day, month, humidity, wind speed, temperature, and pollutant concentrations from stationary stations were included. Adding these predictors significantly improved model performance, highlighting the complex relationships in LUR models for PM 2.5 predictions and offering valuable insights for air quality assessment.
- Open Access
- 23 March 2025
- Published by Elsevier in Transport Policy
- Vol. 167, 66-77
Noise impacts and social justice analysis of off-peak deliveries in the Greater Toronto Area
- 23 March 2025
- Published by Elsevier in Transport Policy
- Vol. 167, 66-77
Open AccessNoise impacts and social justice analysis of off-peak deliveries in the Greater Toronto Area
- Open Access
- 3 April 2023
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2677 (9) , 594-605
Impact of Life Events on Incremental Travel Behavior Change
Travel demand models regenerate the travel behavior of persons or households from scratch at every model run. However, the literature suggests that travel behavior remains relatively stable over time. Change in travel behavior is triggered by life events such as change in employment, household relocation, or birth of a child. The inability of existing travel demand models to represent habitual travel behavior and change in travel behavior of a person/household becuase of life events tends to exaggerate policy sensitivity and result in longer model run times to recreate travel behavior for every agent. In this study, we examined the travel behavior of persons between two consecutive years using a mobility panel survey from Germany. The travel behavior of persons with and without a life event is compared econometrically. Here, the travel behavior is measured as the number of weekly trips by activity type and mode and the impacts of six types of life events are studied. The results show that life events affect travel behavior, but the degree of impact varies by the type of life event, the trip purpose, and the mode. In some cases the impact is found to be negligible, but for many other cases the impact is profound. Moreover, general trends (not affected by life event) in travel behavior are also found. It is concluded that such dynamics in travel behavior should be represented by travel demand models for more sensible policy testing and computationally efficient travel demand models.
- 3 April 2023
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2677 (9) , 594-605
Open AccessImpact of Life Events on Incremental Travel Behavior Change
Travel demand models regenerate the travel behavior of persons or households from scratch at every model run. However, the literature suggests that travel behavior remains relatively stable over time. Change in travel behavior is triggered by life events such as change in employment, household relocation, or birth of a child. The inability of existing travel demand models to represent habitual travel behavior and change in travel behavior of a person/household becuase of life events tends to exaggerate policy sensitivity and result in longer model run times to recreate travel behavior for every agent. In this study, we examined the travel behavior of persons between two consecutive years using a mobility panel survey from Germany. The travel behavior of persons with and without a life event is compared econometrically. Here, the travel behavior is measured as the number of weekly trips by activity type and mode and the impacts of six types of life events are studied. The results show that life events affect travel behavior, but the degree of impact varies by the type of life event, the trip purpose, and the mode. In some cases the impact is found to be negligible, but for many other cases the impact is profound. Moreover, general trends (not affected by life event) in travel behavior are also found. It is concluded that such dynamics in travel behavior should be represented by travel demand models for more sensible policy testing and computationally efficient travel demand models.
- 28 July 2022
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2677 (2) , 1445-1454
GPS-Based Traffic Conditions Classification Using Machine Learning Approaches
This paper addresses the problem of accurately estimating traffic conditions based on sparse GPS information. GPS data have limited spatial-temporal availability, particularly at a regional scale. Therefore, it lacks reliability to accurately estimate traffic conditions of a transportation network. This study proposes a novel methodology to address this problem. First, instead of estimating traffic conditions on a geographic road segment, traffic conditions are estimated for trip segments, which span multiple road segments. Second, machine learning methods are applied to classify traffic conditions. In this study, traffic conditions are defined as the combination of congestion level and road type. This study develops two machine learning models—a random forest (RF) model and a supervised clustering method—to classify traffic conditions, using trip characteristics such as average speed and acceleration. The two models are compared in relation to their accuracy and computational efficiency. Results show that speed-related trip characteristics, such as average instantaneous speed, are the most important variables for classifying traffic conditions in both methods. In addition, the proportion of idling in a trip is essential in distinguishing the Congested Highway and Uncongested Urban traffic conditions when applying the supervised clustering method. The comparison shows that the RF model has a higher estimation accuracy (81%) than the supervised clustering method (72%). Overall, this study shows that traffic conditions can be determined efficiently even in cases of limited GPS data.
- 28 July 2022
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2677 (2) , 1445-1454
GPS-Based Traffic Conditions Classification Using Machine Learning Approaches
This paper addresses the problem of accurately estimating traffic conditions based on sparse GPS information. GPS data have limited spatial-temporal availability, particularly at a regional scale. Therefore, it lacks reliability to accurately estimate traffic conditions of a transportation network. This study proposes a novel methodology to address this problem. First, instead of estimating traffic conditions on a geographic road segment, traffic conditions are estimated for trip segments, which span multiple road segments. Second, machine learning methods are applied to classify traffic conditions. In this study, traffic conditions are defined as the combination of congestion level and road type. This study develops two machine learning models—a random forest (RF) model and a supervised clustering method—to classify traffic conditions, using trip characteristics such as average speed and acceleration. The two models are compared in relation to their accuracy and computational efficiency. Results show that speed-related trip characteristics, such as average instantaneous speed, are the most important variables for classifying traffic conditions in both methods. In addition, the proportion of idling in a trip is essential in distinguishing the Congested Highway and Uncongested Urban traffic conditions when applying the supervised clustering method. The comparison shows that the RF model has a higher estimation accuracy (81%) than the supervised clustering method (72%). Overall, this study shows that traffic conditions can be determined efficiently even in cases of limited GPS data.
- 14 May 2022
- Published by Springer Nature in Transportation
- Vol. 50 (5) , 1613-1629
Joint and sequential models for freight vehicle type and shipment size choice
- 14 May 2022
- Published by Springer Nature in Transportation
- Vol. 50 (5) , 1613-1629
Joint and sequential models for freight vehicle type and shipment size choice
- Open Access
- 23 September 2021
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2676 (2) , 541-552
Modeling Freight Vehicle Type Choice using Machine Learning and Discrete Choice Methods
The choice of vehicle type is one of the important logistics decisions made by firms. The complex nature of the choice process is because of the involvement of multiple agents. This study employs a random forest machine learning algorithm to represent these complex interactions with limited information about shipment transportation. The data are from Commercial Travel Surveys with information about outbound shipment transportation. This study models the choice among four road transport vehicle types: pickup/cube van, single-unit truck, tractor trailer, and passenger car. The characteristics of firms and shipments are used as explanatory variables. SHAP-based variable importance is calculated to interpret the importance of each variable, and shows that employment and weight are the most important variables in determining the choice of vehicle type. The random forest model is also compared with the multinomial and mixed logit models. The model prediction results on the validation data are compared. The results show that random forest model outperforms both the multinomial and mixed logit model with an overall increase in accuracy of about 7.8% and 9.6%, respectively.
- 23 September 2021
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2676 (2) , 541-552
Open AccessModeling Freight Vehicle Type Choice using Machine Learning and Discrete Choice Methods
The choice of vehicle type is one of the important logistics decisions made by firms. The complex nature of the choice process is because of the involvement of multiple agents. This study employs a random forest machine learning algorithm to represent these complex interactions with limited information about shipment transportation. The data are from Commercial Travel Surveys with information about outbound shipment transportation. This study models the choice among four road transport vehicle types: pickup/cube van, single-unit truck, tractor trailer, and passenger car. The characteristics of firms and shipments are used as explanatory variables. SHAP-based variable importance is calculated to interpret the importance of each variable, and shows that employment and weight are the most important variables in determining the choice of vehicle type. The random forest model is also compared with the multinomial and mixed logit models. The model prediction results on the validation data are compared. The results show that random forest model outperforms both the multinomial and mixed logit model with an overall increase in accuracy of about 7.8% and 9.6%, respectively.
- Open Access
- 9 April 2020
- Published by Springer Nature in Transportation
- Vol. 48 (3) , 1481-1502
Microscopic activity sequence generation: a multiple correspondence analysis to explain travel behavior based on socio-demographic person attributes
Activity sequencing is a crucial component of disaggregate modeling approaches. This paper presents a methodology to analyse and predict activity sequence patterns for persons based on their socio-demographic attributes. The model is developed using household travel survey data from Germany. The presented method proposes an efficient approach to replace complex activity-scheduling modules in activity-based models. First, the paper describes a multiple correspondence analysis technique to identify the correlation between activity sequence patterns and socio-demographic attributes. Secondly, a probabilistic model is developed, which could predict likely activity sequence patterns for an agent based on the results of the multiple correspondence analysis. The model is predicting activity sequence patterns fairly accurately. For example, the activity sequence pattern home–work–home is well predicted ( $${\mathrm{R}}^{2}$$ = 0.99) for all the workers, and the activity sequence pattern home–education–home is rather well predicted ( $${\mathrm{R}}^{2}$$ = 0.90) for students. The model predicts the 112 most common activity sequence patterns reasonably well, which covers 72% of all activity sequence patterns observed.- 9 April 2020
- Published by Springer Nature in Transportation
- Vol. 48 (3) , 1481-1502
Open AccessMicroscopic activity sequence generation: a multiple correspondence analysis to explain travel behavior based on socio-demographic person attributes
Activity sequencing is a crucial component of disaggregate modeling approaches. This paper presents a methodology to analyse and predict activity sequence patterns for persons based on their socio-demographic attributes. The model is developed using household travel survey data from Germany. The presented method proposes an efficient approach to replace complex activity-scheduling modules in activity-based models. First, the paper describes a multiple correspondence analysis technique to identify the correlation between activity sequence patterns and socio-demographic attributes. Secondly, a probabilistic model is developed, which could predict likely activity sequence patterns for an agent based on the results of the multiple correspondence analysis. The model is predicting activity sequence patterns fairly accurately. For example, the activity sequence pattern home–work–home is well predicted ( $${\mathrm{R}}^{2}$$ = 0.99) for all the workers, and the activity sequence pattern home–education–home is rather well predicted ( $${\mathrm{R}}^{2}$$ = 0.90) for students. The model predicts the 112 most common activity sequence patterns reasonably well, which covers 72% of all activity sequence patterns observed.