Harnessing GIS and Machine Learning for Accurate Traffic Flow Prediction in Tehran, Iran

The study assessed five distinct machine learning algorithms for traffic level classification: Decision Tree (DT), K-Nearest Neighbors (KNN), discriminant analysis, Naïve Bayes, and artificial neural networks (ANN). The Decision Tree (DT) algorithm uses a hierarchical structure comprising a root node, internal nodes that pose queries, branches that represent possible outcomes, and leaf nodes that yield predictions. KNN is a straightforward, nonparametric regression method prevalent in data mining, machine learning, and pattern recognition, classifying data points by examining their neighbors and often used in forecasting analysis. Discriminant analysis functions as a multivariate statistical approach that differentiates observation groups based on multiple variables, evaluating each variable’s role in that distinction. The Naïve Bayes classifier, based on Bayes’ theorem, formulates predictions for specific classes by presuming variable independence. It utilizes Bayesian theory for forecasting future events based on current data. ANNs are adept at recognizing patterns and classifications in fields such as simulation and pattern recognition, composed of artificial neurons (nodes) and synapses (edges), classified into feed-forward and feedback networks. A particular two-layer feed-forward network featuring hidden sigmoid neurons is employed. The hyperparameter configurations for the algorithms are as follows:

For DT,

Max Number of Splits: 300; Split Criterion: Gini’s Diversity

For KNN,

K = 1; Distance Metric: City Block; Distance Weight: Inverse

For Discriminant Analysis,

Discrimtype: Quadratic

For Naive Bayes,

Kernel type: Gaussian; Support: Unbounded

For ANN,

Number of Hidden Nodes = 30,45,55; Epochs = 1000; Function: scaled conjugate gradient backpropagation

Prior to conducting the analysis, the dataset was divided into 70% for training and 30% for testing. The performance of the models was assessed using both accuracy and the kappa metric. The results indicate the following outcomes: for Decision Trees (DT), the accuracy was 54.73% with a kappa value of 0.4; for K-Nearest Neighbors (KNN), the accuracy reached 96.14% and the kappa value was 0.95; for Discriminant Analysis, the accuracy was 41.57% with a kappa of 0.27; for Naive Bayes, the accuracy stood at 52.14% and the kappa was 0.38; and finally, for Artificial Neural Networks (ANN), the accuracy was 57.76% with a kappa value of 0.45.

These findings suggest that KNN significantly outperformed the other models. The superior effectiveness of KNN can be attributed to its ability to consider neighbouring traffic conditions, rendering it particularly well-suited for predictions that rely on spatial data.

A key aspect of the study was creating a GIS-based traffic prediction map. By combining GIS with machine learning models, a detailed traffic forecast map was produced, showcasing congestion levels on various city streets. The colour-coded traffic map serves as a practical resource for quickly grasping traffic conditions. It employs a colour scheme to represent different levels of traffic density. Green indicates no traffic, perfect for smooth driving. Yellow shows low traffic, implying that while some vehicles may be present, the flow is largely good. Orange denotes medium traffic, where delays might begin but remain manageable. Red highlights high traffic levels, suggesting possible slowdowns and the need for caution. Lastly, dark red signifies very high traffic, often leading to considerable congestion and extended delays. This system enables drivers to swiftly evaluate the situation and adjust their routes as needed. These maps provide commuters with real-time insights, facilitating the selection of less congested paths and effective trip planning.