Probabilistic Spatio-Temporal Data Inference in Sparse Mobile Crowdsensing With Conditional Diffusion Model

Mobile crowdsensing (MCS) is a data sensing paradigm that recruits mobile users with smart devices to collect data. However, the collected spatio-temporal sensing data in real-world scenarios are often incomplete due to participants availability or transmission loss. Sparse MCS has been proposed as a data inference technology to infer values for unsensed regions using sensed data. Existing approaches typically rely on deterministic inference, which struggles to capture the inherent uncertainty in sensing data, and employ autoregressive methods that suffer from error accumulation. To address these issues, we propose a conditional diffusion probabilistic framework named Diffusion-MCS for efficient spatio-temporal sparse MCS data inference. First, it leverages the generative power of diffusion models to explicitly account for sensing data uncertainty and alleviate error accumulation through iterative and non-autoregressive inference refinement. Then, a noise prediction model is designed with a learnable representation module for coarse-grained sensing data interpolation, a conditional feature extraction module that captures spatio-temporal correlations and geographical dependencies as global context features, and a noise estimation module that iteratively refines noisy samples using the extracted features via cross-attention mechanisms. Extensive experiments on four real-world urban sensing datasets demonstrate the effectiveness of the proposed method compared to seven baselines across various sensing ratios and missing patterns.