Wireless Link Scheduling via Graph Representation Learning: A Comparative Study of Different Supervision Levels

Navid NaderializadeharXiv 2021 · OpenReview

TL;DR

This paper investigates how different levels of supervision (supervised, unsupervised, and self-supervised) impact the performance of graph neural networks (GNNs) for wireless link scheduling. They find that while unsupervised methods can be more sample-efficient on smaller networks, supervised learning generally leads to better overall performance, especially in complex, larger networks, albeit at a higher labeling cost.

What Problem Does This Paper Address?

Optimizing wireless communication networks, particularly for tasks like link scheduling (deciding which transmitter-receiver pairs can transmit simultaneously), is crucial for maximizing network throughput. Traditional methods often struggle with the complexity of interference in these networks. Machine learning, specifically graph neural networks (GNNs), offers a promising avenue by representing the network as a graph and learning scheduling policies. However, a key challenge in applying machine learning to such problems is the need for labeled data. Generating optimal labels for link scheduling via exhaustive search becomes computationally prohibitive as the network size increases. This paper addresses the question of how to train effective GNN-based link scheduling models when high-quality labeled data is scarce or expensive, by comparing the trade-offs of using supervised, unsupervised, and self-supervised learning approaches.

Methods

The core of the paper uses a Graph Neural Network (GNN) backbone to process the wireless network's interference graph. This GNN converts the channel matrix into "node embeddings," which represent each transmitter-receiver pair. These embeddings are then fed into a "link scheduling head" to predict optimal power control decisions (effectively, whether to schedule a link). The novelty lies in how these GNNs are trained:

  • Supervised Learning: Trained using ground-truth optimal link scheduling decisions obtained from exhaustive search.
  • Unsupervised Learning: Trained to directly optimize the system's sum-rate objective function without explicit labels.
  • Self-Supervised Learning (SSL): Used as a pre-training step for the GNN backbone. This involves training the GNN to distinguish between augmented versions of the same graph, helping it learn robust representations before fine-tuning with supervised or unsupervised losses.

The paper then empirically compares these approaches across various metrics.

Datasets

The authors generated synthetic datasets for wireless interference channels with varying numbers of transmitter-receiver pairs (K = 4, 6, 8, 10). For each network size, they generated 256 training and 256 testing samples. The channel gains were generated using a dual-slope path-loss model with log-normal shadowing. The paper does not introduce a new benchmark dataset but uses synthetic data representative of typical wireless network scenarios.

Key Insights and Ideas

  • The Labeling Bottleneck is Real: The paper visually and quantitatively demonstrates how the cost of generating optimal labels for link scheduling explodes with network size, motivating the exploration of less supervised methods.
  • Unsupervised Learning Shines in Simplicity: For smaller networks, unsupervised learning can achieve comparable or even better results than supervised learning with less training data, highlighting its potential when labels are scarce.
  • Supervised Learning Dominates Complexity: As network size and interference complexity increase, supervised learning models generally outperform unsupervised ones in terms of throughput, convergence speed, and generalization. This suggests that explicit guidance from optimal solutions is crucial for handling intricate interference patterns.
  • Self-Supervised Pre-training is a Powerful Booster: SSL pre-training significantly accelerates the convergence of both supervised and unsupervised models, demonstrating its value in improving sample efficiency and learning robust initial representations.
  • Generalization Depends on Training Scale: Models trained on smaller networks tend to generalize better to other small networks (unsupervised), while models trained on larger networks generalize better to other large networks (supervised), indicating that the complexity of the training data influences the learned model's adaptability.

Key Results

  • Both supervised and unsupervised models achieved over 80% of the optimal sum-rate from exhaustive search.
  • For smaller networks (K=4), unsupervised models outperformed supervised models with fewer training samples (as few as 64).
  • For larger networks (K=10), supervised models consistently outperformed unsupervised models across all training sample sizes.
  • Self-supervised pre-training reduced convergence time, allowing supervised models to reach a 0.8 normalized sum-rate within the first two epochs for K ∈ {4, 6}.
  • Unsupervised models trained on Ktrain=4 showed better generalization to larger Ktest than supervised models trained on Ktrain=4. Conversely, supervised models trained on Ktrain=10 generalized better to larger Ktest than unsupervised models trained on Ktrain=10.

Limitations and Future Work

  • Limitations: The study relies on synthetic data. The performance of unsupervised learning is heavily dependent on the ability to define and optimize the objective function, which may not always be feasible in real-world scenarios where the objective is unknown or non-differentiable.
  • Future Work: The authors suggest exploring reinforcement learning approaches for scenarios where the objective function is unknown or non-differentiable. They also propose comparing reinforcement learning with supervised learning, particularly for modeling temporal aspects in underlying problems, as an interesting future research direction.

Summary updated

Abstract

We consider the problem of binary power control, or link scheduling, in wireless interference networks, where the power control policy is trained using graph representation learning. We leverage the interference graph of the wireless network as an underlying topology for a graph neural network (GNN) backbone, which converts the channel matrix to a set of node embeddings for all transmitter-receiver pairs. We show how the node embeddings can be trained in several ways, including via supervised, unsupervised, and self-supervised learning, and we compare the impact of different supervision levels on the performance of these methods in terms of the system-level throughput, convergence behavior, sample efficiency, and generalization capability.
BibTeX
@misc{naderializadeh2021wireless,
  author = {Navid Naderializadeh},
  title = {{Wireless Link Scheduling via Graph Representation Learning: A Comparative Study of Different Supervision Levels}},
  year = {2021},
  eprint = {2110.01722},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.01722},
}
Figure