Joint Multidimensional Pattern for Spectrum Prediction Using GNN.

Area of Science:

• Wireless communication
• Signal processing
• Machine learning

Background:

• Efficient spectrum access is crucial for cognitive radio secondary users (SUs).
• Electromagnetic spectrum data from SUs exhibit correlations across time, frequency, and space.
• Multidimensional forecasting of spectrum status is key for efficient resource utilization.

Purpose of the Study:

• To propose a novel deep learning hybrid model, TensorGCN-LSTM, for cognitive radio spectrum sensing.
• To efficiently forecast the spectrum status by leveraging multidimensional correlations in spectrum data.

Main Methods:

• Developed a TensorGCN-LSTM model utilizing a tensor data structure.
• Constructed two graph structures representing spatial and frequency domains of SUs.
• Employed graph convolutional operations for feature extraction and LSTM for fusing spatial, frequency, and temporal features.

Main Results:

• The TensorGCN-LSTM model demonstrated superior prediction performance compared to baseline models (LSTM, GCN, GC-LSTM).
• Achieved a low Root Mean Square Error (RMSE), indicating high prediction accuracy.
• A correlation coefficient R² of 0.8753 confirmed the model's feasibility and effectiveness.

Conclusions:

• The proposed TensorGCN-LSTM model effectively predicts wireless spectrum usage by capturing multidimensional correlations.
• This approach enhances the efficiency of spectrum resource allocation for cognitive radio users.
• The model's performance validates its potential for real-world cognitive radio applications.