TTDI: Transformer Based Truthful Data Inference in Sparse Mobile Crowdsensing

Abstract

In Sparse Mobile Crowdsensing, many data inference schemes have been proposed without considering the sensing dataset with different trust levels which contains unrecognizable false data, resulting in high inference error and poor practical utility. To address this challenge, we propose a Transformer based Truthful Data Inference (TTDI) scheme that can tolerate false data interference to enhance data inference quality. Firstly, the trust-weighted loss function (TLF) is proposed to reduce the false interference for data inference model training. The TLF implements a differential model parameter updating strategy for data with different trust levels, which can enhance and weaken the effect of high-trust and low-trust data on the trained data inference model, respectively. Secondly, the transformer based truthful spatiotemporal feature extraction (T-TSFE) algorithm is proposed to mitigate the mutual interference between biased spatiotemporal features. The T-TSFE extracts data features through temporal and spatial multi-head attention respectively, which reduces the propagation of false information across the spatiotemporal dimensions. Finally, we conduct extensive experiments with four real-world air quality datasets to verify the effectiveness of the TTDI in data inference quality.