Traffic congestion in cities poses a significant challenge, impacting public health, the environment, and the economy. Effective management strategies are critical for mitigating traffic congestion and enhancing flow efficiency. Since supervised machine learning (ML) and unsupervised techniques have been widely applied to traffic flow and congestion prediction, they face notable limitations, such as high computational costs, dependency on high-quality data, limited adaptability to dynamic traffic patterns, and scalability challenges in handling large datasets. These constraints highlight the need for hybrid approaches that integrate clustering and supervised learning to overcome some of these issues. The proposed hybrid approach integrates the concept of K-means clustering with some supervised machine learning techniques, such as Linear Regression (LR), Decision Tree (DT), Gradient Boosting (GB), Random Forest (RF), and Support Vector Regression (SVR) to identify distinct traffic patterns based on features such as road length, traffic volume, and time of day, thereby improving traffic congestion prediction. Our proposed hybrid approach significantly improves predictions compared to individual machine learning techniques, as evidenced by better Root Mean Square Error (RMSE), Coefficient of determination or R-squared (R2), and Mean Absolute Error (MAE) values than those achieved using only supervised machine learning methods. The RMSE for GB decreases from 2180.85 (without clustering) to 620.84 (with clustering), and for RF, it drops from 2186.32 to 632.07. Similarly, in case of DT the RMSE value decreases from 2188.05(without clustering) to 640.80(with clustering). The R2 scores improved significantly when K-means clustering integrated with Gradient Boosting, Random Forest, and Decision Tree, reaching 0.92 compared to lower values without using clustering. SVR also shows some improvements with clustering as well. Its RMSE decreases significantly from 1531.82 (without clustering) to 795.05 (with clustering), and its R2 score improves from 0.065 to 0.713. Additionally, SVR achieves the lowest Mean Absolute Error (MAE) of 301.46, highlighting its effectiveness in minimizing prediction errors after clustering. Conversely, LR performs poorly due to its inability to capture non-linear relationships in traffic data. These results indicate that integrating K-means clustering with ensemble models effectively captures underlying traffic patterns, leading to more accurate predictions.

Analysis of Traffic Flow Congestion by Integrating Supervised Machine Learning with K-mean Clustering

期刊论文
发表日期：

2025年3月6日

查看全文

Sumit Gupta,
Awadhesh Kumar,
Akhilesh Kumar

摘要

Traffic congestion in cities poses a significant challenge, impacting public health, the environment, and the economy. Effective management strategies are critical for mitigating traffic congestion and enhancing flow efficiency. Since supervised machine learning (ML) and unsupervised techniques have been widely applied to traffic flow and congestion prediction, they face notable limitations, such as high computational costs, dependency on high-quality data, limited adaptability to dynamic traffic patterns, and scalability challenges in handling large datasets. These constraints highlight the need for hybrid approaches that integrate clustering and supervised learning to overcome some of these issues. The proposed hybrid approach integrates the concept of K-means clustering with some supervised machine learning techniques, such as Linear Regression (LR), Decision Tree (DT), Gradient Boosting (GB), Random Forest (RF), and Support Vector Regression (SVR) to identify distinct traffic patterns based on features such as road length, traffic volume, and time of day, thereby improving traffic congestion prediction. Our proposed hybrid approach significantly improves predictions compared to individual machine learning techniques, as evidenced by better Root Mean Square Error (RMSE), Coefficient of determination or R-squared (R²), and Mean Absolute Error (MAE) values than those achieved using only supervised machine learning methods. The RMSE for GB decreases from 2180.85 (without clustering) to 620.84 (with clustering), and for RF, it drops from 2186.32 to 632.07. Similarly, in case of DT the RMSE value decreases from 2188.05(without clustering) to 640.80(with clustering). The R² scores improved significantly when K-means clustering integrated with Gradient Boosting, Random Forest, and Decision Tree, reaching 0.92 compared to lower values without using clustering. SVR also shows some improvements with clustering as well. Its RMSE decreases significantly from 1531.82 (without clustering) to 795.05 (with clustering), and its R² score improves from 0.065 to 0.713. Additionally, SVR achieves the lowest Mean Absolute Error (MAE) of 301.46, highlighting its effectiveness in minimizing prediction errors after clustering. Conversely, LR performs poorly due to its inability to capture non-linear relationships in traffic data. These results indicate that integrating K-means clustering with ensemble models effectively captures underlying traffic patterns, leading to more accurate predictions.