Intelligent Systems and Machine Learning for Traffic Prediction on Suburban Roads

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 24 2024

In a paper published in the journal Scientific Reports, researchers employed intelligent systems to analyze vast traffic data on a suburban road in northern Iran, aiming to predict traffic states. They utilized principal component analysis, genetic algorithms, and cyclic features to manage data dimensions effectively.

The models, including long short-term memory (LSTM), support vector machine (SVM), and random forest (RF), demonstrated that integrating cyclic features improved traffic state prediction accuracy compared to a base model with all initial features. The extended short-term memory model, incorporating 71 cyclic features, achieved the highest accuracy while reducing the modeling execution time from 456 to 201 seconds.

Related Work

Previous research has delved into the transformative impact of intelligent transportation systems on data collection methods, particularly emphasizing the removal of constraints associated with traditional approaches reliant on human resources. The focus has been on big traffic data from suburban roads in Iran, characterized by high volume, velocity, and variety.

Numerous studies have demonstrated the superiority of machine learning algorithms over statistical models in analyzing and predicting traffic parameters. However, the abundance of observations and features in big data poses challenges, leading researchers to explore data management techniques.

Comprehensive Methodology for Traffic Prediction

The methodology of this study comprises three key components: data dimension management, machine learning models, and the evaluation of traffic state prediction indexes. The first part explores various data dimension management methods, including principal component analysis (PCA), genetic algorithms (GA), and cyclic features.

PCA, an unsupervised dimensionality reduction technique, constructs relevant variables through linear combinations of initial variables, aiding in identifying essential features explaining total variance. GA is employed to determine influential traffic state features and eliminate irrelevant ones. Cyclic features involve assigning numbers to different time-related aspects, such as season, month, day, and hour, and utilizing trigonometric functions to generate cyclic features. The aim is to enhance or preserve prediction accuracy using fewer predictor features.

The methodology also emphasizes machine learning (ML) models, including LSTM, SVM, and RF, with LSTM leveraging its capability to handle temporal dependencies by utilizing short-term memory, long-term memory, and observations for predicting traffic states. SVM aims to find the best boundary between classes, utilizing the radial basis function kernel for non-linear data. RF, comprising multiple decision trees, provides predictions based on the most repetitive outcomes of individual trees. Each model's hyperparameters and values are defined to suit the study's objectives.

Researchers estimated model performance evaluation using accuracy, F1 score, and modeling execution time. Accuracy and F1 scores are calculated based on test data, while modeling execution time in seconds is an additional criterion for comparing the efficiency of data dimension management algorithms. The study aims to determine the impact of these methods on modeling execution time and evaluate their effectiveness in maintaining or improving prediction accuracy. Confusion matrices visualize true positive and false negative instances, providing insights into the models' performance.

Data Dimension Management Findings

The results of data dimension management methods indicate a substantial reduction in features used as input for the models. Initially, nominal or ordinal features necessitated transformation into dummy variables, resulting in a dataset with 178 features. PCA was then employed to select 30 principal components, explaining 37% of the total variance. GA removed certain features related to holidays, blockages, and opposite-direction blockages. Cyclic features represented time-related aspects, enhancing or preserving prediction accuracy while utilizing fewer predictor features.

The performance of machine learning models, including LSTM, SVM, and RF, was evaluated using different feature sets, namely all features (base), PCs, cyclic features, and GA-selected features. The results reveal that PCs slightly reduce the accuracy of all models compared to the base models with all initial features. GA-selected features maintain accuracy, decreasing by 1 to 2%, while cyclic features increase accuracy by approximately 3 to 4%. LSTM consistently outperforms SVM and RF in accuracy, with the LSTM model utilizing cyclic features achieving the highest accuracy of 88.09%.

State B exhibits higher accuracy, while state C has the least accuracy due to its lower frequency in the datasets. Cyclic features consistently enhance F1 scores for all states and models, with an average increase of 6.5%, 4.4%, and 7.9% for states A, B, and C, respectively. The F1 metric proves valuable for evaluating models with imbalanced datasets.

Modeling execution time reveals that utilizing only a subset of PCs leads to the shortest execution time. Despite reducing prediction accuracy compared to using all features, GA and PCA contribute to shorter modeling execution times. This advantage is particularly significant for online models with processing power and time constraints. The noteworthy finding is that using cyclic features outperforms the base model in modeling execution time and prediction accuracy. RF and SVM models exhibit the minimum and maximum execution times, respectively, with the LSTM model trained with cyclic features ranking fifth in terms of modeling execution time among the 12 models.

Conclusion

In conclusion, this study emphasizes the importance of short-term traffic state prediction for efficient transportation management on suburban roads. The study utilizes large-scale traffic data and machine learning approaches to improve modeling efficiency by applying data dimension management techniques such as PCA, cyclic features, and GA.

While PCA reduces prediction accuracy and GA maintains it, cyclic features prove essential in boosting accuracy. The LSTM model is the most accurate among RF and SVM, particularly when incorporating cyclic features, achieving a prediction accuracy of 88.09%. However, limitations include the unavailability of certain critical variables. Future research avenues could explore alternative data processing methods to refine predictive models further.