Incentive Mechanism Design Based On Worker Dependence For Truth Discovery Of Crowdsourcing

Mobile crowdsourcing is a distributed problem-solving model,in which a crowd of undefined size and an open platform are engaged to solve the complex problems.With the rapid popularization of smart phones,mobile crowdsourcing has become an effective method to solve large-scale applications.It is essential that requesters use data quality and information and the extracted information from platform to provide accurate decisions.In mobile crowdsourcing,the difficulty of the task,as well as the workers'expertise and willingness,can affect data quality.In particular,the workers with different temporal and spatial backgrounds and levels of personal effort may submit data with different accuracy.In addition,the rational workers tend to maximize their utilites strategically when performing tasks,which may reduce data quality.For example,careless data is received in crowdsourcing-based applications.The participants can even perform tasks by copying the answers of others,instead of completing the tasks independently.Such low quality submissions will reduce the quality of crowdsourcing.In order to provide the high-precision data to the requester,this thesis solves the maximum social welfare problem to complete all tasks while meeting the minimum truth-value precision requirement.This thesis designs an incentive mechanism IMC~2(Incentive Mechanism for Crowdsourcing with Copiers),including the truth discovery phase and the reverse auction phase.In the truth discovery phase,this thesis converts dependencies between workers into dependencies between values and estimates the truth value of each task based on the correlation and accuracy between workers and proposes the DATE(Dependence and Accuracy based Truth Estimation).In the reverse auction phase,the worker selection and incentive problem are abstracted into a SOAC(Social Optimization Accuracy Coverage)problem,and the SOAC problem is proved to be an NP-hard problem.A greedy algorithm is designed to select the winner and determine the reward.Filtering some similar values before calculating worker independence can get the worker's accuracy more accurately.Therefore,based on the same system model,this thesis further studies the truth discovery of worker correlation based on semantic analysis.This thesis learns a vector representation of each value of each task based on a corpus of historical tasks.This thesis uses the offline neural network model BERT(Bidirectional Encoder Representations from Transformers)to learn the similarity between information,construct content vectors in the semantic scene and use the adaptive clustering algorithm KANN-DBSCAN(K-Average Nearest Neighbor Density-Based Spatial Clustering of Applications with Noise)to group the obtained content vectors.Through rigorous theoretical analysis and extensive simulations,this thesis proves that IMC~2can achieve computational efficiency,individual rationality,truthfulness and guaranteed approximation.When there are duplicates in the crowdsourcing system,our truth discovery shows significant advantages in terms of accuracy.In addition,the worker correlation method based on semantic analysis can allow workers to upload text data and calculate worker accuracy more accurately.