Incentive Mechanism Design For Supply Side In Data Trading

With the rapid growth of massive data and the strong demand for large-scale data integration and application in various industries,data trading has encountered a rare opportunity for development.The smooth progress of data trading relies on the continuous and stable supply of data from the data supply side.However,most data suppliers are selfish and rational individuals who are unwilling to participate in data trading without appropriate incentives.Therefore,research on incentive mechanisms for the data supply side of data trading is extremely important.Currently,incentive mechanisms are mainly applied to three typical trading scenarios:crowdsourcing(trading data itself),federated learning(trading models trained on local data),and federated crowdsourcing(trading models trained on local data with dynamic data collection).In different trading scenarios,the diverse behavioral strategies of the participants on the supply side present many challenges to the design of incentive mechanisms.In crowdsourcing scenarios,there may be dishonest behavior among participants,and the accurate modeling and management of the reliability of participants will affect the quality of traded data.In federated learning scenarios,there may be competition among participants,and the fairness of the incentive mechanism will affect the willingness of participants to trade.In federated crowdsourcing scenarios,human resources are required for data collection and labeling,and the efficiency of participants may dynamically change.The platform’s management and scheduling need to balance model performance and participants’ rest time.To address these issues,this paper will design reliable,fair,and efficient incentive mechanisms for data trading in the three major scenarios based on the reliability,contribution,and productivity of participating parties.Specifically,the main research content and contributions of this paper include:·This work proposes an incentive mechanism based on the fine-grained reliability of participants for numerical tasks in crowdsourcing scenarios.In crowdsourcing,accurate modeling of participant reliability is a prerequisite for designing incentive mechanisms to obtain high-quality data.However,existing reliability models cannot be applied to the granularity of participant reliability in numerical tasks.Therefore,this work proposes the first fine-grained participant reliability model based on latent topics in crowdsourcing scenarios for numerical tasks.Based on this reliability model,online task allocation algorithms and truth inference algorithms are designed to improve the quality of transaction data.Specifically,this article proposes a fine-grained participant reliability model that mines latent topics in numerical tasks from participants’ historical behaviors and estimates participants’ topic-level reliability.Based on reliability,this work designs two online task allocation mechanisms that dynamically assign the most appropriate task to each participant using entropy changes and potential contributions to improve the quality of data collection.In addition,this work proposes a true value inference algorithm based on EM,which accurately infers the true value and latent topics of numerical tasks and dynamically updates participants’ topic-level reliability based on the deviation between participant answers and inferred true values.The smaller the deviation,the more reliable the participant,thereby incentivizing participants to provide high-quality data through the management of participant reliability.Based on experiments validated on multiple real datasets,this work is able to achieve reliability management while improving the quality of trading data,thereby encouraging participants to join crowdsourcing.·This work proposes a fair incentive mechanism based on participant contributions in the federated learning(FL).In FL,fairness is an important goal in designing incentive mechanisms.However,existing monetary incentive mechanisms ignore differences in participant contributions,resulting in all participants sharing the same FL model.This is unfair to participants who contribute more and can lead to free-rider behavior.To address this issue,this work proposes the first nonmonetary incentive mechanism for FL that satisfies both aggregation fairness and reward fairness.Aggregation fairness is achieved by detecting local gradient quality and aggregating based on data quality.Reward fairness is achieved through a contribution evaluation based on Shapley values and a reward allocation method based on reputation and gradient distribution.Specifically,this work first uses marginal loss to detect quality and filter out low-quality local gradients,and then uses marginal loss to determine aggregation weights and aggregate the model.Next,it introduces a contribution metric based on Shapley values,evaluating the similarity between local gradients and global gradients as a utility function to assess participants’ contributions each round.Finally,this work combines the participants’ reputation(computed by quality detection and contribution calculation)with the distribution of local and global gradients to select the appropriate number of gradients from the global gradient and construct a model to be allocated to each participant,matching the model with their contribution.Based on experiments validated on multiple real datasets,this work is able to achieve a good balance between model performance and fairness,thereby encouraging participants to join federated learning.·This work proposes an efficient incentive mechanism based on participant productivity in federated crowdsourcing scenario.In federated crowdsourcing,human labor is needed for data collection and labeling,and effective labor management is important to encourage more participants to join the transaction.However,existing works mainly focus on maximizing participants’ work output to improve system efficiency and/or federated learning model performance,ignoring the balance between participants’ work and rest.Therefore,this work proposes the first labor management method based on participant productivity in the context of federated crowdsourcing.Specifically,this work first measures each participant’s contribution to the model and estimates the urgency of collecting and labeling new data based on the rate of contribution change.Then,it calculates the working time based on the participant’s maximum productivity and mood.Finally,considering the urgency of obtaining new data and the participant’s productivity,the work provides scheduling services based on Lyapunov optimization,suggesting whether participants should collect and label new data at the beginning of each round.Based on experiments validated on multiple real datasets,this work is able to achieve a good balance between model performance and participants’work time,thereby encouraging participants to join federated crowdsourcing.