Time-series analysis with smoothed Convolutional Neural Network | JoVE Visualize

Area of Science:

Computer Science
Artificial Intelligence
Data Science

Background:

Convolutional Neural Networks (CNNs) are primarily known for image processing, not time-series forecasting.
Time-series forecasting accuracy is heavily dependent on input data quality, often necessitating data preprocessing techniques like smoothing.
Existing forecasting methods may not fully leverage the potential of deep learning architectures for complex time-series patterns.

Purpose of the Study:

To introduce a novel hybrid forecasting model, Smoothed-CNN (S-CNN), integrating exponential smoothing with CNNs.
To evaluate the performance of S-CNN against established forecasting techniques, including standard CNN, Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM).
To demonstrate the efficacy of combining data smoothing with deep learning for enhanced time-series prediction.

Main Methods:

Developed a hybrid model (S-CNN) by incorporating exponential smoothing techniques into a CNN architecture.
Utilized a dataset comprising a year of daily website visitor data for time-series forecasting.
Employed the Lucas numbers to determine the optimal number of hidden layers in the neural network, addressing the lack of predefined rules.
Compared S-CNN performance against baseline CNN, MLP, and LSTM models using a standard train-test split (80%:20%).

Main Results:

The Smoothed-CNN (S-CNN) model demonstrated superior forecasting performance compared to MLP and LSTM.
The hybrid S-CNN approach achieved the best Mean Squared Error (MSE) of 0.012147693.
Optimal performance was observed with 76 hidden layers within the S-CNN architecture.
The combination of smoothing and CNNs proved more effective than individual methods.

Conclusions:

The S-CNN model represents a significant advancement in time-series forecasting by effectively integrating data smoothing with CNNs.
Hybrid approaches combining preprocessing techniques with deep learning models offer superior predictive accuracy.
S-CNN provides a robust and accurate method for forecasting time-series data, particularly for datasets like daily website visitors.