GKNet: Graph Kalman Filtering and Model Inference via Model-based Deep Learning

Mohammad Sabbaqi, Riccardo Taormina, Elvin Isufi · 2025

Read GKNet: Graph Kalman Filtering and Model Inference via Model-based Deep Learning by Mohammad Sabbaqi, Riccardo Taormina, Elvin Isufi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-22004.txt.

Introduction

Data that moves across a network while it changes in time needs a filter that respects both, and a 2025 arXiv paper by Mohammad Sabbaqi, Riccardo Taormina, and Elvin Isufi builds one. Their state space model lets both the hidden state and the readings follow the graph's own wiring, with few parameters to learn.

Noise spreads along the edges, so the state can stray from its smooth path in some neighborhoods. The authors first fit the model by likelihood, clean but limited, then wrap the same structure in a deep network that learns it directly. The target jobs are forecasting and filling gaps in partly watched networks.

Catalog record

CREATORMohammad Sabbaqi, Riccardo Taormina, Elvin Isufi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-22004

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GKNet: Graph Kalman Filtering and Model Inference via Model-based Deep Learning Mohammad Sabbaqi, Riccardo Taormina, Elvin Isufi Abstract Inference tasks with time series over graphs are of importance in applications such as urban water networks, economics, and networked neuroscience. Addressing these tasks typically relies on identifying a computationally affordable model that jointly captures the graph-temporal patterns of the data. In this work, we propose a graph-aware state space model for graph time series, where both the latent state and the observation equation are parametric graph-induced models with a limited number of parameters that need to be learned. More specifically, we consider the state equation to follow a stochastic partial differential equation driven by noise over the graphs’ edges accounting not only for potential edge uncertainties but also for increasing the degrees of freedom in the latter in a tractable manner. The graph structure conditioning of the noise dispersion allows the state variable to deviate from the stochastic process in certain neighborhoods. The observation model is a sampled and graph-filtered version of the state capturing multi-hop neighboring influence. The goal is to learn the parameters in both state and observation models from the partially observed data for downstream tasks such as prediction and imputation. The model is inferred first through a maximum likelihood approach that provides theoretical tractability but is limited in expressivity and scalability. To improve on the latter, we use the state-space formulation to build a principled deep learning architecture that jointly learns the parameters and tracks the state in an end-to-end manner in the spirit of Kalman neural networks. The proposed models are evaluated on controlled and four real-world scenarios, showing the effectiveness of the proposed approach when the model lacks training data. Finally, we conduct a case study on water networks, where the proposed model is used to predict the water levels in the network when the data is extremely scarce. I Introduction Inference of time-varying signals over graphs plays a key role in network-based systems to interpolate missing values, extrapolate a certain horizon, and detect anomalies, to name a few [ kolaczyk_statistical_2014 , ortega_overview_2018 , marisca_learning_2022 , yi_fouriergnn_nodate , li_deep_2023 ] . This is a challenging problem because the relatio-temporal coupling in the data needs to be exploited in a computationally affordable manner due to the problem dimensions involving large graphs and long time horizons. The challenge increases when the graph is imperfect, which is often the case in almost all physical networks (water, power, transport) but also when the said graph is estimated from a finite amount of data [ giannakis_topology_2018 , dong_learning_2019 , mateos_connecting_2019 , buciulea_learning_2022 , shen_kernel_2017 ] . The impact of the topological uncertainties has been studied in different settings including graph filtering [ isufi_graphfilters_2024 ] , graph signal processing [ ceci_graph_2020 ] , graph neural networks [ sabbaqi_gtcnn , gao_learning_2023 , gao_stochastic_2021 , gao_stability_2021 ] , and PDE-based graph machine learning [ chamberlain_grand_2021 , song_robustness_2022 ] . A natural way to infer time-varying signals on graphs is to extend graph kernels [ smola_kernels_2003 ] into a relatio-temporal form, where a separable kernel is favored due to computational aspects [ romero_kernelbased_2017 , lu_probabilistic_2021 ] . Another approach is to design a latent space model where independent Gaussian processes pass through a graph filter to account for both the temporal and the spatial connections in the data [ lu_spatiotemporal_2021 , zhi_gaussian_2023 ] . The work in [ nikitin_non_2022 ] also builds around Gaussian processes but designs a joint (non-separable) kernel based on stochastic partial differential equations (SPDEs). The latter is more interpretable and expressive than the separable kernels, but it is computationally heavier. In general, these statistical approaches are limited in scalability, and rely on statistical assumptions that may not hold. Another approach to model time-varying signals over graphs is to use temporal variants of graph neural networks (GNNs). One stream of works focuses on sequentially combining GNNs with recurrent neural networks or their variants [ jin_survey_2024 ] . Another stream of works, related to this paper, designs graph neural networks that can capture both relational and temporal dependencies simultaneously. For example, the authors in [ ruiz_gated_2020 ] insert graph convolutions within a GRU architecture to consider spatial connections while exploiting temporal patterns. The work in [ sabbaqi_gtcnn ] propose an architecture performing on a product graph of spatial connections and temporal dependencies. The main drawback of these approaches is that they require a large amount of data to learn the model parameters, and treat the graph structure as solely the domain where graph convolutions are performed, ultimately, limiting the interpretability of the process dynamics. To overcome the limitations of the above approaches, one needs to combine the interpretability and reliability of statistical approaches with the flexibility and scalability of machine learning approaches. Model-based deep learning [ shlezinger_model-based_2022 , shlezinger_model-based_2023 ] is a promising direction in this regard, which aims to develop data-driven solutions where particular models or respective optimization problems are used as inductive biases to design principled architectures that allow a degree of tractability [ hadou_robust_2023 ] . Model-based deep learning for graphs has been proposed in [ sabbaqi_gtfunrolling , chen_graph_2021 , nagahama_graph_2022 , chen_time_2021 , vu_Unrolling_2021 ] . However, these models are restricted to a certain task, typically time-invariant, and data. The latent perspective is often key to model the time series evolution as the driving factors may not be observed [ isufi_observing_2020 ] . Moreover, they work either on static graph signals or short temporal windows. State space model-based deep learning solutions are a powerful alternative in this regard. In this context, the authors in [ revach_kalman_2022 ] proposed KalmanNet for tracking state variable which is a deep learning architecture that learns the Kalman gain parameters via a recurrent neural network. Later on, [ sagi_extended_2023 ] proposed a spectral-based approach for extended Kalman filtering in non-linear graph state space models. Based on this work, the GSP-KalmanNet [ buchnik_gspkalman_2023 ] is proposed to track graph time-varying signals in state space models similar to KalmanNet. However, they consider both the observation and transition models known which is rarely the case. In this work we propose a graph-based state space model for inferring time varying signals over graphs where both the state and observation models are unknown. This is a high-dimensional ill-posed problem, and, to make it tractable, we consider the state equation based on a graph SPDE kernel that accounts for the uncertainties in the graph structure. The observation model is a locally graph-filtered version of the latent graph signals to prevent over-smooth predictions. We use said parametric state space formulation to model the process evolution on two fronts. First, both the state and observation models are learned through an optimization problem in a maximum likelihood setting alongside the latent time series. This method is theoretically grounded and interpretable, but it is computationally heavy and lacks expressivity. Thus, we extend the proposed algorithm into a model-based deep learning architecture with a recurrent inference model to enhance its robustness, expressive power, and scalability even in the presence of topological uncertainties and few data samples. Our main contributions are fourth fold: • Graph-based state space model with uncertainties: We propose a graph-based state space model for inferring time-varying signals over graphs that accounts

Read the complete text