Missing Data Inference in Spatio-Temporal Mobile Crowd Sensing Using Graph Neural Network
Year: 2025
Type: conference-paper
Abstract: Mobile crowd sensing (MCS) is a technology where sensors on mobile devices gather multiple types of data. These data can be used in a variety of analytical procedures. However, the restricted spatio-temporal coverage of the obtained data frequently causes performance problems for classical MCS. This has led to the development of Sparse MCS, which uses data inference methods to rebuild entire datasets from sparsely gathered user input. Even with its potential, current Sparse MCS techniques usually ignore gaps or spatiotemporal inconsistencies where data is completely missing. These gaps can be for a particular time slot across all sensing regions (spatial gap) or for a specific sensing area over all time slots (temporal gap). These gaps pose serious difficulties for data inference algorithms because they make it hard to capture complex spatio-temporal connections in the data. To address this issue, we suggest using a graph-based representation to represent the crowd-sensing dataset, because graphs are non-Euclidean, which makes them more effective at handling dynamic datasets. In particular, we built a graph structure that capture the temporal and spatial characteristics of the dataset. To work with the graph-structured dataset, we employ the Graph Neural Network model, which is specifically developed for graphs. That is, we suggest a neural network model that prioritizes both temporal and spatial relationships, based on the Graph Neural Network. Our suggested graph-based neural network model for inference of missing data in mobile crowd-sensing datasets is effective, according to the results of our experiments. We evaluate our model on two real-world datasets: hourly weather data from Brazil and air pollution data from Seoul. Experiments simulate missing data scenarios from 10% to 90% to assess robustness. The results show that our model significantly outperforms traditional imputation techniques, including uniform and weighted K-Nearest Neighbors (KNN). Under extreme conditions (90% missing data), the proposed model achieves a Normalized Mean Absolute Error (NMAE) of approximately 0.3, whereas KNN-based methods show much higher errors, with NMAEs reaching 0.6. These findings confirm the model's effectiveness in handling incomplete and complex MCS datasets, highlighting its effectiveness and potential.
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4414539433