SS-RNN: A Strengthened Skip Algorithm for Data Classification Based on Recurrent Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Recurrent Neural Networks (RNNs) are prevalent for time series tasks but suffer from limited memory and gradient propagation challenges.
• Existing models like Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), and MCNN have limitations in handling long-term dependencies and gradient stability.

Purpose of the Study:

• To propose a novel algorithm, SS-RNN, designed to enhance memory capabilities and gradient propagation in RNNs for time series prediction.
• To investigate the impact of incorporating multiple historical data points and different processing methods on RNN performance.

Main Methods:

• Developed the SS-RNN algorithm, which directly integrates multiple historical time steps for current prediction.
• Introduced two processing methods (continuous and discontinuous) and two historical information addition techniques (direct addition and weighted/mapped addition), creating six distinct pathways for exploration.
• Evaluated SS-RNN against LSTM, Bi-LSTM, GRUs, and MCNN using real-world datasets, analyzing Accuracy, Precision, Recall, and F1-score.

Main Results:

• SS-RNN demonstrated improved average accuracy compared to established models like LSTM, Bi-LSTM, GRUs, and MCNN.
• The proposed methods effectively enhanced the long-term memory capacity and temporal correlation within the RNN structure.
• SS-RNN successfully mitigated the exploding and vanishing gradient problems inherent in traditional RNNs.

Conclusions:

• The SS-RNN algorithm offers a significant advancement in time series prediction by enhancing memory and addressing gradient issues.
• The flexible design of SS-RNN allows for deep exploration of historical data's influence, optimizing RNN performance.
• SS-RNN provides a robust and effective solution for complex time series analysis, outperforming existing state-of-the-art methods.