Task Recommendation Method For Mobile Crowd Sensing Based On Multimodal Information Fusion

With the rapid development of intelligent mobile terminals and wireless sensor networks,mobile crowdsensing systems are being deployed on a large scale,and related applications are increasingly suitable for people ’s daily living environment.As a new perception paradigm emerging in recent years,it takes the user ’s smartphones,tablets,sports watches and other portable mobile devices as the basic perception unit,and uses the built-in sensors of these perception nodes to form a large interactive and participatory perception network.Through the participation of the majority of users to collect the perception data of the surrounding environment,it is a significant embodiment of group wisdom in the field of mobile data perception.Compared with traditional wireless sensor networks,the data collected by mobile crowd sensing systems are more multimodal,time-sensitive and spatio-temporal.As the core module in the research of mobile crowdsensing,task recommendation mainly studies how to effectively use user characteristics to recommend appropriate sensing tasks for users,so as to improve user satisfaction,maximize the quality of perceived data,optimize the perception process,and then make mobile crowdsensing system popular.Therefore,this paper studies the task recommendation method in mobile crowd sensing.The main research results are as follows:1.A mobile crowd sensing task recommendation method based on heterogeneous multimodal features and decision fusion is proposed.First,a task-task similarity matrix is constructed according to the content features of text and image modal data in the user history task set to achieve alignment in multi-modal feature dimension and semantic dimension.Then,the improved similarity network fusion algorithm is used to effectively fuse multiple content similarity networks into a similarity network,and the user ’s preference pattern is updated with the forgetting law to filter out the tasks that have been migrated.Finally,the iteratively updated similarity network is clustered to predict users ’ current preferences for newly published tasks.The experimental results on different datasets show that this method can improve the accuracy and efficiency of task allocation.Moreover,this method can also improve the robustness of the recommendation system.2.A mobile crowd sensing task recommendation method based on multi-view user dynamic behavior prediction is proposed.Firstly,starting from the multi-view behavior sequence,the attention mechanism is used to set different weights for different user individual behaviors according to the strength of social influence,and the aggregation representation of group user behavior at different time granularities is calculated.Then,the memory neural network is used to fuse the single user multi-scale behavior sequence and the group user multi-scale behavior sequence to extract the single user multi-view embedded behavior sequence features.Finally,through multilabel prediction,the user ’s preference probability for multiple perceptual behaviors is obtained,and the user ’s multi-aspect preference for perceptual task types is predicted simultaneously.Combining the experimental results of multiple data sets,it is proved that this method can effectively reduce the sensing cost and deal with the cold start problem compared with other baseline methods.3.A multi-domain collaborative mobile crowd sensing task recommendation method based on variable information bottleneck is proposed.First,through the variational bipartite graph encoder,the homogeneous information of similar neighbors is aggregated,and the respective user / item latent variable representation of the platform is generated.Then,the variable information bottleneck regularizer is used to integrate various representations in multiple domains,reduce the influence of redundant information,and denoise the cross-domain user-item interaction representation from its source domain.Finally,the overlapping user information regularizer is used to refine the representation of overlapping users,further capture the domain sharing information from the two platforms,complete the joint modeling between the two domains,and finally cross-domain task recommendation.Simulation results show that the proposed method outperforms other baseline methods in terms of sensing task coverage and sensing user satisfaction.