Research On Key Technologies Of Spatio-Temporal Data Oriented Sparse Mobile CrowdSensing

Mobile crowdsensing is an emerging and promising paradigm that encourages people to contribute data sensed or generated by their smart mobile devices.Then data is transmitted to the cloud for analysis,management,and processing,enabling the integration of collective intelligence and serving the public.For traditional large-scale urban sensing applications,due to the wide coverage area and long duration of the target region,platforms usually need to recruit many participants to cover most of the sensing areas to obtain fine-grained and high-quality urban sensing maps.However,the high sensing cost(network bandwidth consumption,participant incentives,computing and storage costs,etc.)caused by fine-grained sensing hinders the sustainable development of platform applications.To achieve high-quality sensing with low sensing cost,researchers have started to focus on the ”Sparse mobile crowdsensing” scheme.This scheme recruits only a small number of suitable participants to sense a few(sparse)urban target subareas.Then it utilizes the spatio-temporal correlations among different subareas’ data to infer data of unobserved subareas,significantly reducing the number of sensing subareas and thereby decreasing the overall sensing cost.However,the field of sparse mobile crowdsensing still faces the following challenges.Firstly,the issue of essential subareas may be unsensed due to the uneven spatial and temporal distribution of crowds and the problem of redundant selection of subareas caused by fine-grained and uniform partitioning of sensing areas,leading to decreasing task assignment quality.Secondly,due to the complexity of factors involved in selecting important subareas and the variations in participants’ mobility patterns across different sensing modes,traditional methods struggle to effectively integrate human resources from various sensing modes,thus affecting the acquisition of high-quality sensing maps with lower sensing costs.Thirdly,for the complex scenario of performing federated prediction learning tasks across sparse crowdsensing platforms,it is difficult for each platform to establish accurate and practical federated prediction models due to concerns about data privacy breaches and high communication overhead.To address these challenges,this paper conducted three main research works with the following contributions:(1)In response to the detrimental effect of the problem of the uneven spatial and temporal distribution of crowds and uniform partitioning of sensing areas on the quality of the original task assignment scheme,this paper proposes a subarea division learning based task allocation framework in sparse mobile crowdsensing.For the sensing area division issue,we propose a subarea division learning algorithm,which jointly considers historical data and spatio-temporal correlations to perform uneven subarea division,providing guidance for more reasonable task assignment schemes and more accurate sensing map reconstruction.On this basis,to achieve the goal of minimizing sensing map errors while ensuring the sensing map quality,we design a task assignment algorithm based on mutual information and bootstrapping sampling.The task assignment process iteratively selects the most cost-effective sensing cells and participants,which jointly considers the mutual information of sensing cells,sensing capabilities of participants,and sensing mobility costs.To validate the effectiveness of the proposed framework,advanced task assignment frameworks are tested and evaluated on four real-world datasets.Experimental results demonstrate that the proposed framework outperforms existing methods and can achieve more accurate sensing maps.(2)To effectively integrate the human resources of various sensing modes to improve system efficiency,this paper proposes a task assignment framework based on hybrid sensing modes in sparse mobile crowdsensing,fully leveraging the advantages of each sensing mode and achieving a good trade-off between sensing quality and costs.Considering the characteristics and advantages of opportunistic and participatory sensing modes,we propose a heuristic two-stage search algorithm to jointly recruit suitable participants from various modes to perform tasks in important sensing subareas and provide an effective cost allocation scheme to achieve the highest quality of sensing map under the total cost constraint.Compared to quantitative cost allocation schemes,our algorithm can quickly obtain cost allocation schemes close to the optimal solution.Furthermore,to address the issue of inaccurate future trajectory prediction caused by lacking historical trajectory data for new opportunistic participants,we design a mobility prediction model based on transfer learning,which effectively improves the prediction accuracy and enhances the quality of the hybrid task assignment scheme.The experimental evaluation results on two real trajectory datasets and sensor datasets demonstrate that the proposed framework achieves the best performance in improving sensing map quality.(3)For the complex scenario where cross platforms face issues of data privacy leakage and high communication overhead,we propose a sparse mobile crowdsensing prediction model based on data privacy-preserving.The model utilizes federated learning technology to enable joint training of accurate prediction models without the need for sparse crowdsensing platforms to share their perception data.We propose a communication-efficient hierarchical personalized federated framework based on complex network feature clustering to address the poor prediction performance caused by the data heterogeneity issue and the high communication overhead in cross-platform scenarios.This framework groups crowd-sensing platform servers with similar data distributions in a privacy-preserving manner and utilizes the edge-mediator-cloud architecture of hierarchical model aggregation to train personalized federated prediction models for each platform efficiently.Experimental evaluations conducted on three real-world environmental datasets demonstrate that proposed framework algorithm outperforms existing comparison algorithms regarding personalized performance and communication efficiency.