Informed Graph Learning By Domain Knowledge Injection and Smooth Graph Signal Representation

Keivan Faghih Niresi, Lucas Kuhn, Gaëtan Frusque, Olga Fink

TL;DR

This paper introduces a novel graph learning method that injects domain knowledge into graph signal processing. By combining physical insights from district heating networks with smooth graph signal representations, the method significantly improves graph signal reconstruction for tasks like denoising and imputation, outperforming existing approaches.

What Problem Does This Paper Address?

Graph signal processing (GSP) is a powerful tool for analyzing data on complex networks. A crucial step in GSP is learning the underlying graph structure from data. While methods based on "smooth graph signals" (where neighboring nodes have similar signal values) are state-of-the-art, they often ignore valuable domain-specific knowledge about the network. This can lead to less interpretable and less effective graph inference, especially in specialized applications. The authors tackle this by proposing a way to integrate domain knowledge into the graph learning process, aiming to create more accurate and insightful graph representations.

Methods

The core of the paper is a new graph inference method called Informed Graph Learning (IGL). It builds upon existing GSP techniques that use smooth graph signal representations, which often involve minimizing a Laplacian quadratic form. However, to make graph inference more robust and informative, IGL incorporates domain knowledge in two key ways:

  1. Physics-Inspired Graph Construction: For district heating networks, the authors construct an initial graph based on physical characteristics like pressure drop and temperature variations between nodes. This "physics-inspired" graph (WPI) captures relationships that might not be evident from raw data alone.
  2. Informed Graph Learning Optimization: The IGL method then uses an optimization framework that combines the smooth graph signal assumption with the physics-inspired graph. It introduces a new regularization term that penalizes deviations from the domain-knowledge-based graph (WPI) in specific parts of the network, while still allowing the model to learn connections in areas where domain knowledge is less available. This is achieved by solving a complex optimization problem using a primal-dual splitting algorithm.

Datasets

The authors created a synthetic dataset for district heating networks using the TesPy library. This dataset consists of 8760 samples, with 5000 for training and 3760 for testing. To make the synthetic data more realistic, zero-mean Gaussian noise with a standard deviation of 0.25 was added to the training data. Min-max normalization was applied to pressure and temperature sensor readings.

Key Insights and Ideas

  • Domain knowledge is crucial for effective graph inference: Simply relying on data smoothness can miss important network characteristics, leading to suboptimal performance.
  • Physical processes can be translated into graph structures: By analyzing physical phenomena like pressure and temperature drops, we can construct meaningful initial graph representations.
  • Combining data-driven and knowledge-driven approaches yields superior results: The IGL method's strength lies in its ability to leverage both the inherent smoothness of graph signals and the explicit information provided by domain expertise.
  • District heating networks are complex systems where GSP can be applied: This work pioneers the application of GSP to such networks, opening doors for future research in similar infrastructure systems.

Key Results

The IGL method was evaluated on graph signal reconstruction tasks, specifically denoising and imputation, in the synthetic district heating network dataset.

  • Denoising: IGL achieved the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) among all tested methods, outperforming a pure physics-based graph, a Laplacian-smoothness method, and an adjacency-smoothness method.
  • Imputation: For various sampling densities (p), IGL consistently performed competitively, often outperforming the Adj-Smooth method, especially at higher sampling ratios (p=0.7 and 0.9). The pure physics-based graph showed a significant performance drop when data was sparse.
  • Visual Comparison: A colormap of the absolute difference between IGL and Adj-Smooth adjacency matrices (Figure 2) visually highlights the distinct graph structures learned by each method, suggesting IGL captures more nuanced relationships.

Limitations and Future Work

  • Dataset Limitation: The study relies on a synthetic dataset due to the absence of publicly available real-world data for district heating networks.
  • Future Directions: The authors suggest exploring other applications of their method, such as in power grids where voltage drops can inform graph construction. They also propose assessing the method on other tasks beyond denoising and imputation.

Summary updated

Abstract

Graph signal processing represents an important advancement in the field of data analysis, extending conventional signal processing methodologies to complex networks and thereby facilitating the exploration of informative patterns and structures across various domains. However, acquiring the underlying graphs for specific applications remains a challenging task. While graph inference based on smooth graph signal representation has become one of the state-of-the-art methods, these approaches usually overlook the unique properties of networks, which are generally derived from domain-specific knowledge. Overlooking this information could make the approaches less interpretable and less effective overall. In this study, we propose a new graph inference method that leverages available domain knowledge. The proposed methodology is evaluated on the task of denoising and imputing missing sensor data, utilizing graph signal reconstruction techniques. The results demonstrate that incorporating domain knowledge into the graph inference process can improve graph signal reconstruction in district heating networks. Our code is available at \href{https://github.com/Keiv4n/IGL}{github.com/Keiv4n/IGL}.
BibTeX
@misc{niresi2024informed,
  author = {Keivan Faghih Niresi and Lucas Kuhn and Gaëtan Frusque and Olga Fink},
  title = {{Informed Graph Learning By Domain Knowledge Injection and Smooth Graph Signal Representation}},
  year = {2024},
  eprint = {2406.03898},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.03898},
}
Figure