Truth Discovery for Spatio-Temporal Events from Crowdsourced Data - PowerPoint PPT Presentation

Truth Discovery for Spatio-Temporal Events from Crowdsourced Data Daniel García Ulloa, Li Xiong, Vaidy Sunderam Emory University Logo Image Here Author (Institution) Short Title Date or Conference

Content ● Introduction ● Truth Inference in Spatial Crowdsourcing Graphical model ○ Bayesian estimation ○ Bayesian estimation and Kalman filter ○ Experimental Results ○ ● Conclusion Logo Image Here Author (Institution) Short Title Date or Conference

Introduction Logo Image Here Author (Institution) Short Title Date or Conference

Related Work Truth Inference in Spatial Crowdsourcing Based on the techniques used, the algorithms can be classified in: 1) Direct Computation. Do not model workers or tasks. E.g. Majority Voting, Median 2) Optimization - Develop a function that captures the relation between tasks and workers. - Optimize 3) Iterative methods Logo Image Here Author (Institution) Short Title Date or Conference

Related Work Truth Inference in Spatial Crowdsourcing Zheng, Yudian, et al. "Truth inference in crowdsourcing: is the problem solved?." Proceedings of the VLDB Endowment 10.5 (2017): 541-552. Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Truth Inference in Spatial Crowdsourcing Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Truth Inference in Spatial Crowdsourcing Logo Image Here Author (Institution) Short Title Date or Conference

Related Work ● Truth discovery methods in SC can be classified [12] into: Iterative (e.g. Truthfinder [7]) ○ Optimization-based (e.g. MalVoteCount [13]) ○ Probabilistic graphical models (e.g. Latent Truth Model [23]) ○ although overlaps are possible ● Several methods ([19,20,21]) do not consider mobile data ● Other methods (TSE [15]) do not consider events changing over time Logo Image Here Author (Institution) Short Title Date or Conference

Contributions ● Present dynamic graphical model that describes dependencies between hidden and observed variables. ● Present a Bayesian estimation method based on this model ● Enhance the model by incorporating the event model that explicitly describes correlations, and develop an algorithm that uses the Kalman filter to take advantage of the event model ● Simulations and real-world data show the performance (F1 measure) of the methods Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Graphical Model Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Modelling user reliability (low TP,TN (high TP and FP) Untrustworthy Aggressive high FP,FN) (high TP,TN (high TN,FN) Passive Trustworthy Low FP,FN) Noisy Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Bayesian Estimation Estimated Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Bayesian Estimation and Kalman Filter algorithm Recursively Update Z: Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Bayesian Estimation and Kalman Filter algorithm Recursively Update : Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Bayesian Estimation and Kalman Filter algorithm Recursively Update H: Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Bayesian Estimation and Kalman Filter algorithm Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Bayesian Estimation and Kalman Filter algorithm Hidden Observed Estimated Predicted Updated Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Truth Inference Algorithm: Case Study Results 2/22/15 - 2/23/15 - 2/24/15 - 2/24/15 - 2/26/15 - 2/25/15 - 2/25/15 - Potholes (blue squares) according to the Time series of potholes and waze reports. government website and Waze reports of Each data point represents 4 hours. potholes (red dots) in Boston on 2/23/2015 from 12PM to 4PM Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Truth Inference Algorithm: Case Study Results Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Truth Inference Algorithm: Simulation Results Logo Image Here Author (Institution) Short Title Date or Conference

Truth Discovery Truth Inference Algorithm: Running Times Logo Image Here Author (Institution) Short Title Date or Conference

Conclusion ● Method has wide variety of applications by considering a default state ● Can be generalized to more labels ● Incorporating the event model allows the method to depend less on the observed data and enhances the performance. Logo Image Here Author (Institution) Short Title Date or Conference

References Logo Image Here Author (Institution) Short Title Date or Conference

References Logo Image Here Author (Institution) Short Title Date or Conference

The End Thanks! Logo Image Here Author (Institution) Short Title Date or Conference