GN2DI: A Scalable Graph Neural Network Framework for Spatial Missing Data Imputation in Sensor Networks

This paper presents GN2DI, a Graph Neural Network (GNN)-based framework designed for data imputation in sensor networks. GN2DI utilizes a neural graph construction technique that can be optimized for the target task. The proposed method can handle both samples and sensors that were unavailable during the training phase, enabling inductive capabilities for new sensors and samples. GN2DI employs two modules: the first module learns efficient graph representations for the downstream task by following the gradient signal, while the second module imputes the missing values. Various experimental settings were devised to evaluate GN2DI's performance using real-world sensor datasets from different domains. The results show a 2% to 14% improvement over the best benchmark. Additionally, an experiment was conducted to verify the model's performance with new sensors. Also, our ablation study gives insight into the model's performance. The code is available at: https://github.com/AmEskandari/GN2DI.

I. Introduction

It is crucial to address the issue of missing data, which is a common challenge in data-driven tasks across various fields, including medicine [1], psychology [2], and the Internet of Things (IoT) [3]. By 2030, it is estimated that more than 29 billion IoT devices will be connected [4], with the majority being sensors. These sensors may encounter issues such as link congestion, battery failure, memory failure, or interference, leading to gaps in data transmission. Given the importance of sensor data for various applications, it is imperative to devise methods to handle missing values. In IoT systems, sensors are deployed in a network to monitor parameters like temperature in specific areas. Despite the potential for sensor failure, it is uncommon for all sensors to malfunction simultaneously. Additionally, neighboring sensors that observe similar variables often exhibit similar readings due to the geo-similarity effect. We believe that leveraging the available sensor data and their relationships with unavailable sensors will provide a viable approach to imputing missing values.