PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Neural Networks in IoT: A Survey

Guimin Dong, Mingyue Tang, Zhiyuan Wang, Jiechao Gao, Sikun Guo, Lihua Cai, Robert Gutierrez, Bradford Campbell, Laura E. Barnes, Mehdi Boukhechba

arXiv:2203.15935v2cs.LG

TL;DR

IoT systems generate heterogeneous sensor data while operating through interconnected, resource-constrained devices, making effective modeling and deployment challenging. This survey reviews how GNNs represent IoT sensing relationships, organizes applications and datasets, and discusses future directions. Across numerous IoT tasks, GNNs achieve state-of-the-art or comparable performance, while reliability, security, and computational constraints remain important boundaries.

  • Problem

    Heterogeneous IoT sensing data and complex device relationships challenge existing learning and deployment methods.

  • Method

    The survey develops a unified IoT sensing framework, a taxonomy of GNN modeling approaches, and compilations of applications, public datasets, toolkits, and future research directions.

  • Results

    GNNs capture complex relationships and interdependencies in IoT sensor networks and achieve state-of-the-art or comparable performance across numerous IoT tasks.

  • Takeaways & Limitations

    GNN research in IoT spans multi-agent interaction, human state dynamics, and sensor interconnection across diverse sensing applications.

  • Takeaways & Limitations

    IoT GNN deployment remains constrained by communication failures, security threats, and limited device memory, computation, and energy.

Abstract

from arXiv · show

The Internet of Things (IoT) boom has revolutionized almost every corner of people's daily lives: healthcare, home, transportation, manufacturing, supply chain, and so on. With the recent development of sensor and communication technologies, IoT devices including smart wearables, cameras, smartwatches, and autonomous vehicles can accurately measure and perceive their surrounding environment. Continuous sensing generates massive amounts of data and presents challenges for machine learning. Deep learning models (e.g., convolution neural networks and recurrent neural networks) have been extensively employed in solving IoT tasks by learning patterns from multi-modal sensory data. Graph Neural Networks (GNNs), an emerging and fast-growing family of neural network models, can capture complex interactions within sensor topology and have been demonstrated to achieve state-of-the-art results in numerous IoT learning tasks. In this survey, we present a comprehensive review of recent advances in the application of GNNs to the IoT field, including a deep dive analysis of GNN design in various IoT sensing environments, an overarching list of public data and source code from the collected publications, and future research directions. To keep track of newly published works, we collect representative papers and their open-source implementations and create a Github repository at https://github.com/GuiminDong/GNN4IoT.

1 Introduction

IoT systems connect sensing, processing, communication, and mechanical components to support collaborative sensing across increasingly complex environments. As heterogeneous sensor data challenge conventional and deep learning approaches, GNNs model device relationships and interdependencies, motivating a comprehensive survey of GNN-based IoT sensing.

  • IoT devices combine sensing, processing, communication, and mechanical units, using wired or wireless networks to achieve users’ common goals.
  • Connected IoT devices perceive environments and collaborate with other devices and users on complex tasks.
  • Heterogeneous IoT systems make handcrafted feature engineering and conventional analytic methods increasingly challenging.
  • CNNs and RNNs have been widely applied to IoT sensing, including smart health monitoring and electrical load forecasting.
  • GNNs represent complex relationships and interdependencies among IoT devices and temporal data instances, achieving comparable or better performance than existing benchmarks across many applications.
  • The survey unifies IoT sensing into human, autonomous-thing, and environmental paradigms, then develops a GNN taxonomy, reviews applications, compiles public resources, and discusses future directions.

2 Previous Works

Prior surveys have examined GNN theories, taxonomies, and applications in specific domains, but none systematically focuses on GNN applications in IoT sensing. This survey addresses that gap by connecting GNN modeling with networked IoT solutions.

  • Existing review papers cover GNN theories, methodologies, and real-world applications, but few systematically summarize GNNs for IoT sensing.
  • GNN surveys have proposed taxonomies spanning recurrent, convolutional, autoencoder, spatial-temporal, reinforcement-learning, and adversarial paradigms.
  • Other reviews study GNN design through explainability, dynamics, expressive power, accuracy, flexibility, scalability, and efficiency.
  • Application-focused surveys address domains such as recommendation and traffic forecasting, but no survey had focused on GNN applications in IoT.
  • This work fills the gap by examining how GNNs model problems using networked IoT solutions.

3.1 IoT Sensing Objects and Paradigms

IoT sensing spans autonomous things, humans, and environments, while practical tasks often involve interactions among these categories. The survey organizes representative sensing applications and data sources across these overlapping paradigms.

  • IoT sensing paradigms characterize autonomous things, humans, and environments, whose interactions are often captured together rather than separately.
  • Autonomous Things: Autonomous-thing sensing monitors internal operation and external interactions among robots, vehicles, industrial machines, and their environments.
  • Autonomous Things: Connected autonomous vehicle networks aggregate vehicle interactions to support traffic management at the network level.
  • Human: Human-centric sensing unobtrusively captures physiology, behavior, and mobility through ambient sensors and mobile devices.
  • Human: Human IoT applications include healthcare monitoring, social sensing, location-based services, recommendations, and analyses of mobility and social relationships.
  • Environment: Environmental IoT sensing covers urban and natural settings, including traffic monitoring, meteorology, air quality, soil moisture, and other environmental factors.

3.2 Preliminary of Graph Neural Networks

GNNs extend neural processing to graph-structured, non-Euclidean data by representing entities and relationships as nodes and edges. Their core operation aggregates neighborhood information to produce high-level node or graph representations, with several architectural and signal-processing foundations.

  • Graph Neural Networks: GNNs process graph-structured information that conventional CNNs and RNNs cannot directly handle with regular convolutions or recurrent connections.Graphs represent connected structures using nodes and edges, while GNNs map graph inputs into numerical spaces.
  • Graph Notations and Types: A graph is represented as G = {V, E}, with nodes V, edges E, an adjacency matrix A, and a diagonal degree matrix D.The adjacency matrix records node connectivity, while D encodes node degrees.
  • Graph Notations and Types: Graph types differ by edge direction, edge weights, and node or edge heterogeneity, which determine how messages propagate and how graph structure is encoded.Directed graphs constrain message flow, weighted graphs assign edge values, and heterogeneous graphs contain multiple node or edge types.
  • Graph Signal Processing: Graph signal processing adapts Fourier transforms, filters, and convolutions to high-dimensional non-Euclidean IoT sensing data.Spectral-domain operations simplify graph convolutions into multiplications and provide foundations for spectral GNN methods.
  • Graph Neural Networks: GNNs iteratively aggregate neighborhood information and integrate it into node representations through neural information propagation.Architectures are distinguished by their aggregation and node-update operations, including recurrent and convolutional graph neural networks.

4 Categorization of Graph Neural Networks in IoT Sensing

The survey categorizes GNN-based IoT sensing models by how sensing objects and their relationships are represented as graphs. Its three categories are multi-agent interaction, human behavior dynamics, and IoT sensor interconnection, supporting architecture choices tied to domain semantics.

  • Categorization of GNNs in IoT Sensing: The survey divides GNN-based IoT sensing models into multi-agent interaction, human behavior dynamics, and IoT sensor interconnection.The categories are based on the semantic interpretation of graph representations across autonomous things, human, and environmental sensing.
  • Categorization of GNNs in IoT Sensing: Graph nodes and edges have different meanings across sensing scenarios, such as agents and interactive relations in multi-agent systems or human states in behavior modeling.Domain knowledge about interactions determines information transitions within the sensor topology.
  • Multi-Agent Interaction: The multi-agent interaction framework has three steps: perceiving the environment, modeling agent interactions as a graph, and encoding neighboring-agent features with GNNs.Nodes represent agents and edges encode interactive relationships between agents.

4.1 Multi-agent Interaction

The survey frames multi-agent interaction as a problem of representing dynamic relationships among agents in rich sensor environments. Its generalized GNN workflow combines environmental perception, graph modeling, and neighborhood-based representation learning.

  • Motivation: RNN-based models cannot fully exploit dynamic interactive relationships or changing inter-agent spatial topologies.
  • Graph Modeling: GNNs represent intelligent entities as nodes and adjacent entities as edges, then aggregate neighborhood information to encode complex interactions.
  • Framework: The generalized multi-agent interaction framework has three steps: perception of environment, graph modeling in MAI, and GNNs for MAI.
  • Graph Modeling: Graph nodes and edges acquire scenario-specific meanings, such as agents or vehicles as nodes and dynamic road distances as edges.
  • Graph Modeling: Attention mechanisms assign stronger edge weights to more influential neighbors, including closer agents.
  • GNN Kernels: The surveyed GNN kernels include message-passing, graph-convolutional, and graph-attention-based approaches, with message passing potentially facing scalability issues.

4.2 Human State Dynamics

Human-centric sensing produces heterogeneous, multimodal data describing complex and evolving human states. The survey organizes GNN-based human behavior modeling around sensing, graph representation, and behavior inference.

  • Motivation: Human-centric sensor data are multimodal and multichannel, requiring deep learning methods that can fuse heterogeneous sources.
  • GNN Approach: GNNs model human behavior by making interactions among behaviors explicit through graph topologies over sensory data.
  • Framework: The generalized human-state-dynamics workflow consists of human-centric sensing, graph representation of human behaviors or sensors, and behavior inference.
  • Sensing: Human-centric sensing systems include body sensor networks, rich-sensor mobile devices, and combinations of both.
  • Graph Modeling: Graph modeling represents human-state transitions and sensor topology, allowing information to propagate through edges between related nodes.
  • GNN Inference: GNN graph convolutions aggregate multi-hop state information into embeddings that capture interactions among state transitions and connected sensors.

4.3 IoT Sensor Interconnection (ISI)

IoT sensor interconnection modeling addresses environments where diverse sensors must cooperate to capture spatial dynamics and latent relationships. The survey distinguishes geospatial and implicit graph connections and reviews GNNs for these structures.

  • Applications: Sensor interconnections support applications including smart buildings, agriculture, traffic networks, and smart grids.
  • Motivation: Complex IoT applications require multiple sensor infrastructures because a single sensor cannot capture complete geospatial and latent interconnection information.
  • Motivation: Multimodal sensing can reveal causes that individual sensor streams cannot identify, such as relating CO2 changes to motion-detected human activity.
  • Graph Modeling: The survey divides IoT sensor-interconnection graphs into geospatial networks and implicit interconnections learned from latent sensor relationships.
  • GNN Approach: GNNs capture sensor interactions and data-stream correlations that conventional CNNs and RNNs do not provide through graph representations.
  • GNN Types: Spatio-temporal GNNs combine static graph structures with time-varying features to model IoT sensor interconnections.

5.1 Public Datasets and Toolkits

The survey organizes public IoT datasets by their suitability for GNN modeling, distinguishing native graph data, transformable sensory data, and extension sources. It also highlights domain-specific dataset coverage and gaps.

  • Dataset categories: Public datasets are categorized as graph/network structured, non-graph sensory, or extension data for GNN modeling.Non-graph sensory data require preprocessing to form graph structures, while extension data provide node or edge attributes.
  • Graph/network structured data: Air-quality datasets naturally form graphs because measurements come from geographically distributed stations.They commonly include PM2.5, PM10, NO2, CO, SO2, and O3 measurements.
  • Non-graph sensory data: Fall-detection datasets model skeleton joints as graph nodes and acceleration as node attributes.TST V2 and FallFree are identified as computer-vision-based datasets for this setting.
  • Dataset gaps: No public datasets were identified for GNN-based energy forecasting, although energy networks have clear graph structure and application potential.Photovoltaic systems are given as an example of an energy network suitable for GNN-related models.
  • Non-graph sensory data: PSG datasets support sleep-quality prediction, with MASS-SS3 containing 20 EEG, 2 EOG, 3 EMG, and 1 ECG channel per recording.The survey notes that sleep-stage graphs can be constructed after feature extraction.
  • Extension data: Extension sources add domain or expert information to graph data, including weather, radar, and physical atmospheric-model observations.CMA China weather data and NCEP multi-source observations are cited as examples for air-quality prediction.

5.2 GNN Applications in IoT

The survey reviews GNN applications across robotics, healthcare, environmental science, transportation, manufacturing, energy, and smart-home settings. These studies use graph representations to model interactions, sensor relationships, spatial structure, or temporal dependencies.

  • Robotics and autonomous vehicles: GNNs are widely applied to robotics and autonomous-vehicle planning and control because message passing leverages interactions and relational information.The surveyed applications address increasingly complex robotic systems and autonomous-vehicle problems.
  • Health inference and informatics: Healthcare sensing applications include fluid-intake monitoring, sleep-stage detection, structural-health monitoring, fall detection, and human-activity recognition.The survey reports state-of-the-art results for some wearable- and smartphone-based activity-graph approaches.
  • Health inference and informatics: Healthcare has relatively little IoT-and-GNN research because it lacks natural graph scenarios compared with transportation networks and robotics swarms.The survey calls for better graph construction for healthcare problems.
  • Environment: Environmental applications address problems including particulate-matter prediction, water-resource availability, and endocrine-disruptor analysis.The section motivates more advanced analytical approaches as environmental data continue to grow.
  • Transportation: Transportation systems generate large sensor datasets from cameras, GPS, and loop detectors for traffic prediction and control.The survey frames intelligent transportation systems as data-driven responses to increasing transportation complexity.
  • Transportation: OD-demand studies use alternative graph designs, including regions as nodes with OD pairs as edges or OD pairs as nodes linked by demand-pattern similarity.These designs aim to preserve information between origin-destination pairs and can model dynamic node topologies.
  • Manufacturing, energy, and smart home: Manufacturing, energy, and smart-home studies use GNNs for thermal-response prediction, process estimation, photovoltaic reliability, and household-object arrangement.Examples include unseen-geometry thermal histories, dense 3D point clouds, photovoltaic performance, and fully connected object graphs.

6 Challenges and Future Directions

The survey identifies reliability, privacy, security, and computational and energy efficiency as major challenges for deploying GNNs in IoT. It highlights failures, sensitive data exposure, cyber threats, resource constraints, and knowledge distillation as future directions.

  • 6.1 Reliability: Reliability requires IoT neural networks to function in real-world situations despite sensor, communication, and data failures.Such failures can cause inaccurate decisions in smart-city traffic management and healthcare applications.
  • 6.2 Privacy Preservation: Centralized cloud storage can expose sensitive information, including users’ locations and inferred personal attributes from GPS trajectories.Anonymization, federated learning, and differential privacy are identified as privacy-preservation techniques.
  • 6.3 Security: IoT security is constrained because many underlying Internet technologies and communication protocols were not designed for IoT.Threats include exposure of personally identifiable information and malware spreading between connected devices.
  • 6.4 Computational and Energy Efficiency: Resource-constrained IoT devices face prohibitive computational, energy, and storage requirements when implementing deep neural networks.Knowledge distillation offers a future direction by using compressed student models with similar prediction power, though its integration with IoT GNNs remains underexplored.

7 Conclusion

IoT devices have expanded across everyday domains, creating sensor-rich environments with complex multimodal data. The survey reviews how GNNs model sensor topology and organizes existing work around multi-agent interaction, human behavior dynamics, and IoT sensor interconnection.

  • 7 Conclusion: IoT devices connect, cooperate, and collaborate across many domains of everyday life.Their usage has exploded in popularity over the past decade.
  • 7 Conclusion: GNNs capture complex relationships and interdependencies within IoT sensor networks while achieving state-of-the-art results in numerous IoT tasks.The approach considers sensor topology alongside multimodal sensory data.
  • 7 Conclusion: The survey categorizes IoT GNN research into graph modeling of multi-agent interaction, human behavior dynamics, and IoT sensor interconnection.For each category, it reviews sensor infrastructures, general modeling frameworks, and graph components such as nodes and edges.
Loading 2203.15935v2…