Large-Scale Spectrum Occupancy Learning via Tensor Decomposition and LSTM Networks

Mohsen Joneidi, Ismail Alkhouri, Nazanin RahnavardarXiv 2019 · arXiv

Introduction

In the evolving landscape of wireless communication, Spectrum Occupancy Learning (SOL) has emerged as a critical capability for Dynamic Spectrum Access. By predicting when and where frequency bands will be occupied, secondary users can utilize "white spaces" without interfering with primary users. However, traditional SOL methods struggle with the high dimensionality of spectrum data, complex multi-scale temporal correlations, and the frequent occurrence of missing data due to sensor imperfections.

The Challenges of Spectrum Prediction

Predicting spectrum occupancy at scale involves overcoming three primary hurdles:

  1. Multi-Scale Correlations: Spectrum usage exhibits both short-term patterns (minute-by-minute fluctuations) and long-term periodicities (daily and weekly cycles). Standard 1D time-series models often fail to capture these nested relationships.
  2. Computational Complexity: Training deep learning models like Long Short-Term Memory (LSTM) networks on raw, high-resolution spectrum data is computationally intensive, often making real-time application impossible.
  3. Data Incompleteness: Real-world sensing is rarely perfect. Gaps in data can significantly degrade the performance of conventional predictive models.

Proposed Framework: Decompose-then-Predict

To address these challenges, researchers have proposed a novel framework that combines Tensor CP Decomposition with LSTM networks. The core innovation lies in shifting the prediction task from raw data to the underlying low-rank factors of the spectrum data.

1. Tensor Representation

Instead of treating spectrum data as a flat sequence, the framework organizes it into a 3D tensor with dimensions: Frequency (Channel) × Time-of-Day (Short-time) × Day (Long-time). This structure naturally preserves the multi-dimensional correlations inherent in the data.

The 3D tensor representation captures multi-scale temporal patterns by stacking daily measurements across multiple channels.

2. CP Decomposition and LSTM Prediction

The framework employs CANDECOMP/PARFAC (CP) decomposition to break the large 3D tensor into a set of low-dimensional factor vectors. These factors represent the "essence" of the spectrum usage patterns.

The prediction process follows a specific pipeline:

  • Decomposition: The given tensor ($X_L$) is decomposed into spatial factors ($A$), short-term temporal factors ($B$), and long-term temporal factors ($C_L$).
  • Prediction: An LSTM is trained specifically to predict future long-term temporal factors ($C_P$) based on the historical factors ($C_L$). Because these factors are significantly smaller than the raw data, training is orders of magnitude faster.
  • Reconstruction: Finally, the predicted factors are combined with the stable spatial and short-term factors to reconstruct the predicted spectrum occupancy ($X_P$).
The overall framework uses CPD to extract factors, an LSTM to predict future temporal factors, and CPR to reconstruct the final occupancy map.

Experimental Results

The framework was evaluated using a large-scale synthetic dataset mimicking real-world complexities, including 20 channels sampled over 100 days with varying usage peaks.

  • Accuracy: The proposed LSTM+CPD method achieved a Normalized Prediction Error of 15.26%, outperforming plain LSTM models (23.71%) and traditional Auto-Regressive (AR) models (33.55%).
  • Efficiency: The computational speedup was dramatic. The framework processed the data in 15.7 seconds, compared to 2,390 seconds for a standard LSTM approach—a reduction of over 150x.
  • Robustness: The model remained highly effective even with missing data. With 10% of the entries missing, the error rate only increased marginally to 16.53%, demonstrating the inherent denoising and completion capabilities of tensor decomposition.

Conclusion and Future Work

By predicting low-rank factors rather than raw data, this framework provides a scalable and robust solution for large-scale spectrum occupancy learning. While the results on synthetic data are promising, future research will focus on adapting the model for non-stationary real-world radio environments and developing automated methods for selecting the optimal tensor rank, which currently remains a heuristic process.

Summary updated

BibTeX
@misc{joneidi2019large,
  author = {Mohsen Joneidi and Ismail Alkhouri and Nazanin Rahnavard},
  title = {{Large-Scale Spectrum Occupancy Learning via Tensor Decomposition and LSTM Networks}},
  year = {2019},
  eprint = {1905.04392},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.04392},
}
Figure