Creative Commons

This dissertation was written as a part of the MSc in Data Science at the International Hellenic University. Traffic Prediction is an intelligent scheme of forecasting the traffic flow of a specific place. It is the most critical part of any traffic management system in a smart city. Accurate prediction could decrease accidents and time waste and even increase the quality of life of the citizens. That is why; the research of this topic is of the essence. In this thesis, a dataset with traffic flow of 6 different Crosses of unknown place is used with Machine Learning and Deep Learning models. Thus, in order to predict the traffic flow regression models as Linear Regression, Random Forest, Multi Layer Perceptron (MLP) and Gradient Boosting are utilized. Other techniques of analyzing the data were adding “time” features and taking another time interval between the observations of the time series, which concluded to better results. Furthermore, the regression problem has been converted into a classification problem and classifiers such as K-Nearest Neighbors (KNN), Support Vector Classifier (SVC), Adaptive Boosting (Adaboost), Decision Tree, Random Forest, Gaussian Naïve Bayes Classifier (GaussianNB) and Extra Trees are used for experimentation. Last, Long short-term memory (LSTM), that the literature review suggests as one of the top deep learning models to predict traffic flow, was utilized and tuned for our case. Indeed, LSTM outperformed the other models with regards to RMSE metric. At each analysis the according statistical metrics have been calculated to compare the different models and choose the optimal one. In our case, for regression as mentioned the LSTM model was the best one and for classification the Extra Trees and the Random Forest classifiers. Cross Validation and Grid Search had also used in search of optimal models. For the regression problem, a technique that is utilized is that the machine learning models used the data not only of one Cross but of another highly correlated Cross.That results to better models with regards to 𝑅� 2 metric. Thus, different kind of approaches are examined for this univariable type of problem and acquired better results than the classic regression problem.