| Recently,Mobile CrowdSensing(MCS)is a new crowdsourcing paradigm which allocates spatial-temporal sensing tasks to be executed by a large number of participants.The application field of MCS has been very extensive,and it is also a research hotspot in these years.The development of MCS has important societal significance,and its extension of various applications promotes a significant change in current day people's lifestyle.As an important method to realize the all-round sensing perception of human life,the new scientific problems brought about by it have become an important research topic in academia.Meanwhile,it has truly evolved from small-scale applications on campus to all aspects of human life.Through summarizing and investigating the existing research work on MCS,there are many related works on task allocation,data trustworthiness evaluation and privacy protection,which indicated that its related technical route has been relatively comprehensive.However,MCS as a representative of emerging technologies,there are still numerous problems and challenges yet to be solved.This thesis analyzes the key issues in MCS,summarize the current research status and introduce the classical research work with regarding to each stage of MCS's life cycle.Based on the above research,detailed studies have been pursued on location prediction,incentive mechanism and maximal frequent itemset mining,and propose solutions respectively.This thesis includes analysis of basic problems,also the in-depth exploration of specific application scenarios.At the same time,a distributed scheme is presented for large data processing of location prediction and crowdsensing data mining.The main contributions of this thesis include the following three aspects:1.Location Prediction: For participant-task matching and incentive mechanism depending on location prediction in mobile crowdsensing,more accurate location prediction can consolidate the support for more efficient participant recruitment and more reasonable resource allocation.Most previous location prediction studies have tended to mine participants' social relationships in the real world to enhance the mobility model of participants' location prediction.However,this particular data requirement makes it difficult for them to migrate to other platforms.Because these methods focus on social relations,they often ignore the potential value of association between stranger trajectories in group data.Based on this argument,a new partial position prediction method,which is based on trajectory similarity,is proposed in this thesis.The new method uses covering algorithm to accelerate the compression of trajectory sampling,and proposes a trajectory-based participant similarity calculation method through social contagion theory.Computing framework MapReduce is adopted dealing with massive trajectory data processing.Experiments using real trajectory data sets show that the position prediction based on trajectory similarity proposed in this chapter achieves higher accuracy and stability.2.Incentive Mechanism: For most of the existing incentive mechanisms,research tends to auction based incentive mechanism,resulting in the reduction of user participation,marginalization and other issues.In this thesis,a dynamic incentive mechanism based on integrated decision-making system is proposed and applied to the indiscriminately parked shareable bike problem scenario to mobilize people to relocate bikes.The proposed incentive mechanism is based on centralized pricing and its core decision-making problem is formalized into a Multi-dimensional Multiple-choice Knapsack Problem model(MMKP)for overall decision-making and a Cost-refundable,Multiple Resources Constrained Multiple Armed Bandit model(CMRC-MAB)for individual decision-making.A comprehensive solution GA-WSLS is proposed.This algorithm can learn personalized incentives at different times of the day to maximize the overall task utility(the number bikes),while optimizing the usage of multi-dimensional constrained resources in the problem scenarios(budget and number of query).The experiment is based on real-world datasets from Singapore and Brussels.The results show that proposed GA-WSLS outperforms seven classic and state-of-theart methods.3.Maximal Frequent Itemset Mining: Aiming at the problem of balance between data compression and performance in the research of big data processing under the MCS,many works focus on maximal frequent itemset to refine and compress the sensing dataset.Many of these research applications analyze massive sensing data directly through existing mining methods,which leads to problems such as excessive computation,high time complexity,and large storage space.This thesis proposes a Heuristic MapReduce-based Association rule approach through Maximal frequent itemset mining,HMAM.The main idea is: at first,by directly operating the sensing data's transaction database,the transaction is allocated to different processing nodes and all transactions are grouped according to the dimension.Then,Bitmap-Sort is used to screen the most frequent transactions from each transaction set and obtain besttransaction-set by aggregating all transaction-elects of each transaction set.According to the inclusion relationship between transactions in the best-transaction-set,the current candidate maximal frequent item set can be obtained by deleting the sub-transaction.At the same time,each subset of child transactions in the candidate maximal frequent item set is discarded from all transaction sets.Then,the minimum support threshold is used to filter transactions.The final candidate maximal frequent itemset is obtained by iteration until each transaction set is empty.Finally,the maximum frequent itemset of each round are combined to achieve the final acquisition of the maximal frequent itemset.The experimental results show that compared with the existing methods,HMAM significantly avoids a large number of candidate projects generated by the connection operation,speeds up the mining of the maximum frequent itemset,and improves the resource utilization. |
PDF Full Text Request