Beijing Key Laboratory of Internet Culture and Digital Dissemination Research, and School of Computer Science, Beijing Information Science and Technology University, Beijing100101, China
School of Computer Science, Beijing Information Science and Technology University, Beijing100101, China
Show Author Information
Hide Author Information
Abstract
With the rapid development of mobile devices, the use of Mobile Crowd Sensing (MCS) mode has become popular to complete more intelligent and complex sensing tasks. However, large-scale data collection may reduce the quality of sensed data. Thus, quality control is a key problem in MCS. With the emergence of the federated learning framework, the number of complex intelligent calculations that can be completed on mobile devices has increased. In this study, we formulate a quality-aware user recruitment problem as an optimization problem. We predict the quality of sensed data from different users by analyzing the correlation between data and context information through federated learning. Furthermore, the lightweight neural network model located on mobile terminals is used. Based on the prediction of sensed quality, we develop a user recruitment algorithm that runs on the cloud platform through terminal-cloud collaboration. The performance of the proposed method is evaluated through simulations. Results show that compared with existing algorithms, i.e., Random Adaptive Greedy algorithm for User Recruitment (RAGUR) and Context-Aware Tasks Allocation (CATA), the proposed method improves the quality of sensed data by 23.5 and 38.8, respectively.
No abstract is available for this article. Click the button above to view the PDF directly.
References
[1]
D.Estrin, K. M.Chandy, R. M.Young, L.Smarr, A.Odlyzko, D.Clark, V.Reding, T.Ishida, S.Sharma, V. G.Cerf, et al., Participatory sensing: Applications and architecture, IEEE Internet Computing, vol. 14, no. 1, pp. 12-42, 2010.
H. B.McMahan, E.Moore, D.Ramage, S.Hampson, and B. A.Arcas, Communication-efficient learning of deep networks from decentralized data, in Proc. 20th Int. Conf. Artificial Intelligence and Statistics, Fort Lauderdale, FL, USA, 2017, pp. 1273-1282.
[3]
M.Tomasoni, A.Capponi, C.Fiandrino, D.Kliazovich, F.Granelli, and P.Bouvry, Profiling energy efficiency of mobile crowdsensing data collection frameworks for smart city applications, in Proc. 2018 6th IEEE Int. Conf. Mobile Cloud Computing, Services, and Engineering, Bamberg, Germany, 2018, pp. 1-8.
S. Z.Liu, Z. Z.Zheng, F.Wu, S. J.Tang, and G. H.Chen, Context-aware data quality estimation in mobile crowdsensing, presented at IEEE INFOCOM 2017-IEEE Conf. Computer Communications, Atlanta, GA, USA, 2017, pp. 1-9.
A.Hassani, P. D.Haghighi, and P. P.Jayaraman, Context-aware recruitment scheme for opportunistic mobile crowdsensing, in Proc. IEEE 21st Int. Conf. Parallel and Distributed Systems, Melbourne, Australia, 2015, pp. 266-273.
M.Lease, On quality control and machine learning in crowdsourcing, in Proc. 11th AAAI Conf. Human Computation, California, CA, USA, 2011, pp. 97-102.
[7]
C.Fiandrino, F.Anjomshoa, B.Kantarci, D.Kliazovich, P.Bouvry, and J. N.Matthews, Sociability-driven framework for data acquisition in mobile crowdsensing over fog computing platforms for smart cities, IEEE Transactions on Sustainable Computing, vol. 2, no. 4, pp. 345-358, 2017.
C. S.Meng, W. J.Jiang, Y. L.Li, J.Gao, L.Su, H.Ding, and Y.Cheng, Truth discovery on crowd sensing of correlated entities, in Proc. 13th ACM. Conf. Embedded Networked Sensor Systems, Seoul, Republic of Korea, 2015, pp. 169-182.
G. R.Li, S. C.Peng, C.Wang, J. W.Niu, and Y.Yuan, An energy-efficient data collection scheme using denoising autoencoder in wireless sensor networks, Tsinghua Science and Technology, vol. 24, no. 1, pp. 86-96, 2019.
S.Yang, F.Wu, S. J.Tang, X. F.Gao, B.Yang, and G. H.Chen, On designing data quality-aware truth estimation and surplus sharing method for mobile crowdsensing, IEEE Journal on Selected Areas in Communications, vol. 35, no. 4, pp. 832-847, 2017.
X. X.Yin, J. W.Han, and P. S.Yu, Truth discovery with multiple conflicting information providers on the web, IEEE Transactions on Knowledge & Data Engineering, vol. 20, no. 6, pp. 796-808, 2008.
Q.Li, Y. L.Li, J.Gao, L.Su, B.Zhao, M.Demirbas, W.Fan, and J. W.Han, A confidence-aware approach for truth discovery on long-tail data, Proceedings of the VLDB Endowment, vol. 8, no. 4, pp. 425-436, 2014.
X. L.Dong, L.Berti-Equille, and D.Srivastava, Truth discovery and copying detection in a dynamic world, Proceedings of the VLDB Endowment, vol. 2, no. 1, pp. 562-573, 2009.
B.Zhao, B. I. P.Rubinstein, J.Gemmell, and J. W.Han, A Bayesian approach to discovering truth from conflicting sources for data integration, Proceedings of the VLDB Endowment, vol. 5, no. 6, pp. 550-561, 2012.
F.Ribeiro, D.Florêncio, C.Zhang, and M.Seltzer, CROWDMOS: An approach for crowdsourcing mean opinion score studies, in Proc. 2011 IEEE Int. Conf. Acoustics, Speech and Signal Processing, Prague, Czech Republic, 2011, pp. 2416-2419.
Y. Q.Zheng, G. B.Shen, L. Q.Li, C. S.Zhao, M.Li, and F.Zhao, Travi-Navi: Self-deployable indoor navigation system, IEEE/ACM Transactions on Networking, vol. 25, no. 5, pp. 2655-2669, 2017.
S.Reddy, A.Parker, J.Hyman, J.Burke, D.Estrin, and M.Hansen, Image browsing, processing, and clustering for participatory sensing: Lessons from a DietSense prototype, in Proc. 4th Workshop on Embedded Networked Sensors, Cork, Ireland, 2007, pp. 13-17.
E. T. H.Chu, C. Y.Lin, P. H.Tsai, and J. W. S.Liu, Participant selection for crowdsourcing disaster information, in Proc. 3rd Int. Conf. Disaster Management and Human Health Disaster Management, Vienna, Austria, 2013, pp. 231-240.
M.Tomasoni, A.Capponi, C.Fiandrino, D.Kliazovich, F.Granelli, and P.Bouvry, Profiling energy efficiency of mobile crowdsensing data collection frameworks for smart city applications, in Proc. 2018 6th IEEE Int. Conf. Mobile Cloud Computing, Services, and Engineering, Bamberg, Germany, 2018, pp. 1-8.
D. Q.Zhang, H. Y.Xiong, L. Y.Wang, and G. L.Chen, CrowdRecruiter: Selecting participants for piggyback crowdsensing under probabilistic coverage constraint, in Proc. 2014 ACM Int. Joint Conf. Pervasive and Ubiquitous Computing, Seattle, WA, USA, 2014, pp. 703-714.
H.Ko, S.Pack, and V. C. M.Leung, Coverage-guaranteed and energy-efficient participant selection strategy in mobile crowdsensing, IEEE Internet of Things Journal, vol. 6, no. 2, pp. 3202-3211, 2019.
R.Azzam, R.Mizouni, H.Otrok, A.Ouali, and S.Singh, GRS: A group-based recruitment system for mobile crowd sensing, Journal of Network & Computer Applications, vol. 72, pp. 38-50, 2016.
Y. L.Zhang, H. Y.Qin, B.Li, J.Wang, S.Lee, and Z. Q.Huang, Truthful mechanism for crowdsourcing task assignment, Tsinghua Science and Technology, vol. 23, no. 6, pp. 645-659, 2018.
T.Li, A. K.Sahu, A.Talwalkar, and V.Smith, Federated learning: Challenges, methods, and future directions, IEEE Signal Processing Magazine, vol. 37, no. 3, pp. 50-60, 2020.
L. N.Liu, X.Chen, Z. M.Lu, L. H.Wang, and X. M.Wen, Mobile-edge computing framework with data compression for wireless network in energy internet, Tsinghua Science and Technology, vol. 24, no. 3, pp. 271-280, 2019.
X. F.Wang, Y. W.Han, C. Y.Wang, Q. Y.Zhao, X.Chen, and M.Chen, In-edge AI: Intelligentizing mobile edge computing, caching and communication by federated learning, IEEE Network, vol. 33, no. 5, pp. 156-165, 2019.
Q.Yang, Y.Liu, T. J.Chen, and Y. X.Tong, Federated machine learning: Concept and applications, ACM Transactions on Intelligent Systems and Technology, vol. 10, no. 2, p. 12, 2019.
B.Liu, S. J.Tang, X. G.Sun, Q. Y.Chen, J. X.Cao, J. Z.Luo, and S. S.Zhao, Context-aware social media user sentiment analysis, Tsinghua Science and Technology, vol. 25, no. 4, pp. 528-541, 2020.
T.Haruhiko, K.Hidehiko, and H.Terumine, A study on the simple penalty term to the error function from the viewpoint of fault tolerant training, in Proc. 2004 IEEE Int. Joint Conf. Neural Networks, Budapest, Hungary, 2004, pp. 1045-1050.
Zhang W, Li Z, Chen X. Quality-Aware User Recruitment Based on Federated Learning in Mobile Crowd Sensing. Tsinghua Science and Technology, 2021, 26(6): 869-877. https://doi.org/10.26599/TST.2020.9010046
The articles published in this open access journal are distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/).
10.26599/TST.2020.9010046.F001
Federated learning-based user recruitment in MCS.
4.2 Prediction model for sensing data quality
In internet of things networks, wearable devices, autonomous vehicles, or smart homes may contain numerous sensors that allow them to collect large amounts of data in real time for sensing some special scenarios and events[
25
]. However, building analysis models in these scenarios and events may be difficult due to the private nature of data and the limited connectivity of devices. Liu et al.[
26
] proposed a novel framework that utilizes the local area network to collect data from users and reduce transmission latency without extra energy consumption overhead. Some research works have attempted to use new intelligent methods to build the model. Wang et al.[
27
] designed the “In-Edge AI” framework to utilize collaboration among devices and edge nodes intelligently and thus exchange the learning parameters for the improved training and inference of the models. Yang et al.[
28
] proposed building data networks among organizations on the basis of federated mechanisms as an effective solution to allow knowledge to be shared without compromising user privacy.
In this section, we use the federated learning framework to build the prediction model of sensing quality. The system architecture is shown in
Fig. 1
. We divide it into two stages. Firstly, a lightweight neural network is used to construct the relationship between context and sensing data quality. Afterward, a federated learning framework is adopted to solve the relationship model and predict the sensing data quality of mobile users. We predict the data quality of user in sensing task through context information . Context is any information that can be used to model the situation of a specific user[
29
]. This scenario has many context features, such as user identity, time, location, and their activity. Federated learning serves as an on-device distributed training system, which learns a shared global model from distributed mobile devices while keeping the training data at each device. This novel distributed learning paradigm harnesses the benefits of low latency and low power consumption while preserving user privacy. Federated learning gathers mobile users to participate in server-side model training, and the optimization model with constant parameter and variable parameter is as follows:
where is the sensing data quality calculated by the above algorithm.
By training model parameters, the relationship between context and sensing data quality is established. The standard Back Propagation (BP) algorithm uses the squared error as the objective function, and the gradient descent method is used to minimize the error function. As the number of iterations increases, the speed of function approximations decreases. Hence, the accuracy of approximation for highly nonlinear samples cannot be guaranteed. Considering the different relevance of contextual fields to the sensing data quality, we introduce a regularized penalty function to filter the contextual data with low relevance quickly.
Equation (
5
) was proposed in Ref. [
30
] named penalty function to prevent the possible overfitting of error terms, which leads to a decrease in the generalization ability of the network, thereby improving the fault tolerance of the model. This kind of lightweight neural network is built locally at the mobile terminal using the federated gradient descent algorithm for model training. An unequal number of datasets is located on the mobile terminal for user . Meanwhile, the loss function of the local training model is . Federated learning uses an effective parameter aggregation algorithm to train the model. Mobile users can update all model training parameters without uploading local data, thus greatly saving the overhead of the entire perception system. The typical gradient descent algorithm calculates the global parameters as with each round . Based on a fixed learning rate , the model weights can be updated in Formula (
6
):
Meanwhile, the average federated gradient descent algorithm calculates on each user’s local dataset in Formular (
7
), and the central cloud server aggregates the calculation results of each user with a data size of to perform average gradient descent in Formular (
7
).
The local parameter for each user is updated as follows: . The parameter updating algorithm for federated learning model is summarized as Algorithm 1.
4.3 User recruitment algorithm
On the basis of federated learning, we can analyze the relationship between context and sensing data quality and predict the sensing data quality of the user on the basis of the user’s context. Using the predicted sensing quality of different users, we can recruit appropriate users for the sensing tasks. To minimize the sensing cost, we consider the physical location of the users. We design a Quality-Aware User Recruitment (QAUR) algorithm to solve and decide which users are recruited. For every user, we decide whether to recruit it or not. If the answer is yes, we update the current recruited group and the group’s sensed data quality with the predicted results; moreover, the travel cost constraint is updated. If the answer is no, the current state remains as the previous optimal solution. The detailed recruitment process is shown in Algorithm 2.
Dynamic programming is used to solve the problem. In each step, is the maximum sensed quality when recruiting sensing users with a number of with the total distance and the optimal user set for recruitment. Each candidate user has two choice: One is selecting the user to participate in the sensed task and the other is not. We describe the concepts used in the recursion process. On one hand, in the subset without selecting the user , the sensed quality of the optimal subset is . On the other hand, in the subset of selecting the user , the optimal subset is the composition of this user and the optimal subset of , and the sensed quality of this optimal subset is . is the real sensed quality of user . Therefore, the optimal solution for is equal to the greater of these two values. If user is not recruited, then the sensing quality of the best subset is equal to the sensing quality of the best subset selected from the previous user set. Thus, the following recurrence is generated as
In the solution process, we use user context information to predict the user’s current sensing quality to approximate the optimal solution of the problem. The classical dynamic programming approach works from the bottom up, filling the table with solutions to all the small problems. Each of them requires one solution, and some solutions to the smaller problems are not required to solve a given problem; thus, we naturally combine the advantages of the top-down and bottom-up approach. The necessary subproblems are solved only once, and this memory function is used to recruit users every time. The user recruitment algorithm combined with federated learning can not only maximize system resources but also effectively use user context information to estimate its future sensing quality.
10.26599/TST.2020.9010046.F002
Sensing quality of data with different distance constraints.
10.26599/TST.2020.9010046.F003
Sensing quality of data with different numbers of users.