A comparison between machine and deep learning models on high stationarity data.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Big data analytics are increasingly reliant on time series data from advanced sensors and communication technologies.
• Conventional forecasting models often struggle with identifying complex sequence features and achieving high prediction accuracy.
• The need for robust methods to analyze and forecast time series data is critical across various domains.

Purpose of the Study:

• To investigate time series features and compare the performance of machine learning algorithms against deep learning models.
• To analyze the specific problem of predicting vehicle counts at an Italian tollbooth in 2021.
• To determine which algorithms provide superior accuracy for this particular time series forecasting task.

Main Methods:

• Utilized a dataset of 8766 rows and 6 columns related to Italian tollbooth vehicle counts in 2021.
• Applied machine learning algorithms: Support Vector Machine, Random Forest, and eXtreme Gradient Boosting (XGBoost).
• Employed deep learning models: Recurrent Neural Networks with Long Short-Term Memory (RNN-LSTM) cells.
• Evaluated models based on Mean Absolute Error (MAE) and Mean Squared Error (MSE).

Main Results:

• The eXtreme Gradient Boosting (XGBoost) algorithm demonstrated superior prediction performance compared to other machine learning and deep learning models.
• XGBoost achieved lower MAE and MSE values, indicating higher accuracy in forecasting vehicle counts.
• The study found that a shallower machine learning model (XGBoost) adapted better to the time series characteristics than a deeper RNN-LSTM model.

Conclusions:

• Machine learning algorithms, specifically XGBoost, can be more effective than deep learning models like RNN-LSTM for certain time series forecasting problems.
• The choice of algorithm should consider the specific characteristics of the time series data, such as stationarity.
• Simpler, shallower models may offer better adaptation and prediction accuracy in scenarios where deep models tend to over-smooth predictions.