Research On Prediction-based Task Assignment And Adjustment Methods In Spatiotemporal Crowdsourcing

With the increasing popularity of intelligent mobile devices and the rapid development of wireless network technology,spatial crowdsourcing,as a new crowdsourcing model,came into being.In spatial crowdsourcing,task assignment is the fundamental research,which directly related to platform revenue and user experience.The traditional spatial-temporal task assignment methods are usually based on the available participants and tasks at the current time,and rely on the clear mobility patten of participants to assign tasks.However,the traditional task assignment methods can only obtain the local optimal solution.To improve the assignment quality,prediction-based spatial crowdsourcing task assignment method has attracted more and more attention.These methods match the participants and tasks by considering the current state and the future state at the same time,so as to improve the global assignment quality.However,the current prediction-based task assignment methods still face the following issues.First,it is often extremely difficult to accurately predict the future state.Second,for some real-world complex scenes that are difficult to obtain dynamic trajectories in real time,the current methods are difficult to predict the future mobility pattern due to the lack of trajectory data.Third,due to forecast deviation,unexpected situations such as the absence of reserved participants reduce the reliability of the original assignment.In order to solve the above issues,this dissertation includes three main work,and the contributions are as follows:(1)To improve the prediction accuracy,we propose a location-and-preference joint prediction for task assignment,including a multi-attribute joint prediction model and a multi-attribute joint task assignment algorithm based on greedy strategy.Moreover,we analyze the time complexity and theoretical approximation rate of our method.To verify the effectiveness of the proposed method,we evaluate the method on several real GPS trajectory data sets and check-in data sets.The results show that the joint prediction model improves the accuracy of location and preference prediction at the same time.When the task and participant constraints are changed,the average number of completed tasks can be increased by more than 40%.(2)For some spatial crowdsourcing platforms that are difficult to collect participant trajectories,we propose a graph neural network based task assignment method,which uses historical matching information to assist prediction.Through inductive learning,the matching patterns of participants and tasks are mined based on historical matching information.Moreover,we design a greedy-based task assignment algorithm to maximize the matching degree.The results show that the matching degree prediction model improves at different metrics,and the temporal and spatial attributes of the task are beneficial to prediction.When we change the participant constraints,comparing with baselines,the model obtain a better assignment set and improve the user experience.(3)Due to the long prediction time span,the reserved participants may be absent due to temporary events or prediction deviation,which will affect the quality of the original assignment scheme.To break the assignment invariance criterion,we propose a dynamic reassignment method to ensure the reliability of the assignment strategy for the unexpected events.Moreover,we design a multi-objective optimization algorithm to minimize the adjustment and maximize the total matching degree in the reassignment process.The experimental results on three real data sets show that the proposed method can improve the total matching degree by about 10%while minimizing the adjustment compared with the comparison algorithms.