PhD Defense Talk - Near-Optimal Mobile Crowdsensing: Design Framework and Algorithms

78
Institut Mines-Télécom Near-Optimal Mobile Crowdsensing: Design Framework and Algorithms PhD Student: Haoyi Xiong Director of Thesis: Prof. Monique Becker Advisors: Dr. Daqing Zhang, Dr. Vincent Gauthier 22 Jan 2015 xhyccc@gmail .com, http: //fr.linkedin.com/in/haoyixiong/ 1

Transcript of PhD Defense Talk - Near-Optimal Mobile Crowdsensing: Design Framework and Algorithms

Institut Mines-Télécom

Near-Optimal Mobile Crowdsensing: Design Framework and Algorithms

PhD Student: Haoyi Xiong

Director of Thesis: Prof. Monique Becker

Advisors: Dr. Daqing Zhang,

Dr. Vincent Gauthier

22 Jan [email protected], http://fr.linkedin.com/in/haoyixiong/

1

Institut Mines-Télécom

Outline

■ Introduction

• Motivation

• Background & State-of-the Art

■ Technical Contribution

• EEMC

• EMC3

• CrowdRecruiter

• CrowdTasker

■ Conclusion

• Summary

• Future Work

22 Jan 20152

Institut Mines-Télécom

Outline

■ Introduction

• Motivation

• Background & State-of-the Art

■ Technical Contribution

• EEMC

• EMC3

• CrowdRecruiter

• CrowdTasker

■ Conclusion

• Summary

• Future Work

22 Jan 20153

Institut Mines-Télécom

Large-scale Air Pollution Monitoring

World-wide Air Pollution Crisis

• 7—8 million deaths a year (WHO statistics, 2012)

• Increasing risk of lung/bladder cancers

• Even worse in developing countries…

22 Jan 20154

Institut Mines-Télécom

Air Pollution Monitoring using

Traditional Sensor Network

■ Need to deploy expensive

sensors and network

22 Jan 20155

■ Few sensors are deployed

■ Many areas are not covered

16 sensors deployed in Ile-de-France area*

*http://www.eea.europa.eu/themes/air/air-quality/map/real-time-map

Institut Mines-Télécom

Mobile Phone-based Sensing

GPS Sensor

Temperature Sensor

Air Quality Sensor

Sensors

Sensing Application

Location Tracking

Environment Sensing…….

22 Jan 20156

Institut Mines-Télécom

Mobile Crowdsensing (MCS)

Crowd Sensing

Collecting Sensed Results and locations from Mobile Users

Executing Mobile sensing on each Mobile Phone

Fine-grained Air Quality Map

(e.g., UrbanAir, MSRA)

22 Jan 20157

Institut Mines-Télécom

Background and State of the Art

■ Two Major MCS Players

■ Four Steps of MCS Process

■ Five (Common) Research Issues

22 Jan 20158

Institut Mines-Télécom

Two Major MCS Players [Zhang et al.’14]

■ MCS Participants

• The mobile users receiving/performing sensing tasks, and

returning sensed results.

■ MCS Organizers

• The entity that recruit participants for MCS tasks, assign MCS

task to each participant, and collect sensed results.

*Zhang et al. 4W1H of Mobile Crowdsensing, IEEE Communication Magazine

22 Jan 20159

Institut Mines-Télécom

Four Steps of MCS Process

22 Jan 201510

Institut Mines-Télécom

Five Research Issues [Ganti et al. 2011]

■ From MCS Participants Perspectives

• Energy consumption caused by e.g., MCS data transfer, computing,

and sensing on mobile phones

• Individual Incentive Payment e.g., money paid for each user’s

participation

• Privacy e.g., protecting user’s location information (not included in this

thesis)

■ From MCS Organizers Perspectives

• MCS Data Quality e.g., accuracy, coverage of sensor readings

• Total incentive payment e.g., total money paid for all user’s

participation

*Ganti, et al. "Mobile crowdsensing: current state and future challenges." Communications Magazine, IEEE 49.11 (2011): 32-39.

22 Jan 201511

Institut Mines-Télécom

Outline

■ Introduction

• Motivation & Background

• State-of-the Art

■ Technical Contribution

• EEMC

• EMC3

• CrowdTasker

• CrowdRecruiter

■ Conclusion

• Summary

• Future Works

22 Jan 201512

Institut Mines-Télécom

Our Four Technical Contributions

■ With above research issues in mind

• EEMC

− Enabling Energy-efficient Mobile Crowdsensing with Anonymous Participants

(Energy + Data Quality)

• EMC3

− Energy-efficient Data Transfer for Mobile Crowdsensing under Full Coverage

Constraint. (Energy + Data Quality)

• CrowdRecruiter

− Selecting Participants for Piggyback CrowdSensing Under Probabilistic

Coverage Constraint. (Energy + Incentive + Data Quality)

• CrowdTasker

− Allocating MCS task to Participants in order to Maximize Coverage Quality

Under Budget Constraint (Energy + Incentive + Data Quality)

22 Jan 201513

Institut Mines-Télécom

EEMC Research Outline

■ Motivation & Assumption

■ Research Problems

■ Technical Challenges

■ Framework and Algorithms

■ Evaluation and Summary

22 Jan 201514

Institut Mines-Télécom

EEMC (Energy-Efficient Mobile Crowdsensing)

Motivation and Assumption

■ Motivations

• Reducing Individual Energy Consumption

− Two-way Piggyback Crowdsensing using call opportunities

75% energy reduction in MCS data transfer, Numerinen et al. 2010

*Nurminen, Jukka K. "Parallel connections and their effect on the battery consumption of a mobile phone.“,CCNC, 2010. IEEE, 2010.

22 Jan 201515

Institut Mines-Télécom

EEMC Motivation and Assumption

■ Motivations• Reducing Individual Energy Consumption

− Two-way Piggyback Crowdsensing using call opportunities

• Minimizing #task assignments in order to:

− Reduce Total incentive payment, while

− Meeting MCS Data quality requirement

■ Assumptions• Individual Incentive Mechanism

− Pay per task assignment

• MCS Data Quality Requirement

− Dividing MCS process sensing cycles (e.g., two hours)

− Collecting at least a predefined number of sensed from the target region every sensing cycle

22 Jan 201516

Institut Mines-Télécom

EEMC Example

■ Please note:

• future calls are not known in advance.

• Only accumulated Call Traces are Accessible.

− CBD—central business district

22 Jan 201517

Institut Mines-Télécom

EEMC Technical Challenges

– Online Next-Call-Prediction

• Predicting if the new-arriving caller/callee would place

another call in the current sensing cycle, using

accumulated call traces

– Pause or Continue?—Pace Control

• Deciding if tasks already assigned could ensure the

expected number of returning…

– Current User or Future users?—Optimal Task

Assignment Decision Making

• Deciding if assigning the task to current caller/callee or left

the task for future callers/callees

22 Jan 201518

Institut Mines-Télécom

Framework of EEMC

A new phone call comes

Predicting next calls

If a further task assignment is needed?

Current user or future user?

22 Jan 201519

Institut Mines-Télécom

■ Next-n-Call Probability

• Probability of user i placing n calls from time t of cycle k to the

end of cycle k

■ Already-Assigned-Fulfilling Probability

• P{Xk,t(Ak-Rk)≥Ne-|Rk|}—probability of participants already assigned

returning (at least) the expected number of results

■ Future-surer-Fulfilling Probability

• P{X*k,t(FSui⋃(Ak-Rk)≥Ne-|Rk|}—probability of users having not

placed calls but having higher probability of placing two calls in

the future of the cycle fulfilling the task

Core Algorithm of EEMC

22 Jan 201520

Institut Mines-Télécom

Dataset and Evaluation Setups

22 Jan 201521

Statistics of D4D Call Traces for EvaluationEvaluation Region in Cote d'Ivoire

From D4D Data Set

Two Baseline Algorithms• Greedy

• Keep assigning tasks to new calling users, until an expected number of participants have returned their sensed results

• Pace• Consisting of the fist two steps of EEMC• Keep assigning tasks to new calling users, until pace

control decides to stop assigning any new tasks

Institut Mines-Télécom

Results and Comparison

22 Jan 201522

Ne: the expected number of sensor readings

EEMC vs• Pace- 6%--23% fewer task assignment

• Greedy- 27%--63% fewer task assignment

Institut Mines-Télécom

Summary of EEMC Contribution

Individual Energy IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per task assignment

#result/cycle

Task Assignment Minimize Incentives under Specific MCS Data Quality Constraint

Individual Task Execution

Two-way Piggyback Sensing using Calls

Data Collection and Aggregation

Nothing here

Individual Participants’ concerns.. MCS Organizer’ concerns..

22 Jan 201523

Institut Mines-Télécom

Open Issues of EEMC

• MCS Data Quality:

− Data may be collected from some dense area only

− No coverage guarantee

• Individual Incentives:

− Using other (e.g., pay per participant) or multiple (e.g., both

pay per task/participant) incentive payment mechanisms

22 Jan 201524

Institut Mines-Télécom

EMC3 Research Outline

■ Research Overview

■ Research Problems

■ Technical Challenges

■ Framework and Algorithms

■ Evaluation and Summary

22 Jan 201525

Institut Mines-Télécom

EMC3 Research Overview

Individual Energy IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per task assignment

#result/cycleFull coverage

Task Assignment Minimize Incentives under Specific MCS Data Quality Constraint

Individual Task Execution

Two-way Piggyback Sensing using Calls

Data Collection and Aggregation

Nothing here

Individual Participants’ concerns.. MCS Organizer’ concerns..

Beyond EEMC

22 Jan 201526

Institut Mines-Télécom

EMC3 Research Problem

– Input• Given the target region as a set of cell towers,

• Given a series of sensing cycles,

• Given the accumulated call traces and mobility traces of all participants;

– Problem• When a participant places a call in the target region, Deciding if we need to

assign a task to the participant, in order to:

• minimize the total number of task assignments while

• Ensure a given number of participants returning their sensed results in each sensing cycle AND each cell tower in the target region being covered by at least one participant. (full coverage constraint.)

22 Jan 201527

Institut Mines-Télécom

EMC3 Research Challenges (Beyond

EEMC)

■ Online Mobility Prediction

• Given a new-arriving user, predicting in which cell tower the

user will place next calls

■ Pause or Continue? Coverage-based Pace Control

• Given participants already assigned, Predicting if they will fully

cover the target region

■ Current user or Future users? Coverage-based Optimal

Task Assignment Decision Making

• Given participants already assigned and users having higher

probability placing two calls in future, predicting if they can

fully cover the target region

22 Jan 201528

Institut Mines-Télécom

EMC3 Framework and Algorithms

■ Mobility Prediction

• The probability of user i being in cell tower cj

when he/she placing a call

■ Coverage Prediction

• The probability of cell tower cl being covered by

user ui. (cassign is the cell tower of cl being

assigned with the task)

■ Covering Probability

• The probability of cell tower cl being covered by

users in Ak-Rk i.e., the participants already

assigned but having not yet returned

22 Jan 201529

EMC3 Framework

Institut Mines-Télécom

Dataset and Evaluation Setups

22 Jan 201530

CBD Region

Residential Region

Statistics of CBD Call Trace

Statistics of Residential Call Trace

Institut Mines-Télécom

Results and Comparison (CBD Traces)

22 Jan 201531

Number of Task Assignment and Returned Results using CDB Call Traces Ne: the expected number of returned sensed results

Coverage using CDB Call Traces (on each cell tower)

Institut Mines-Télécom

Results and Comparison (Residential

Traces)

22 Jan 201532

Number of task assignments and Returned Results

Coverage of Returned Results (on each cell tower)

Average response time and estimated maximal throughput

Institut Mines-Télécom

Summary and Open Issues of EMC3

• MCS Data Quality:

− Unnecessary to cover all cell towers in the target region

(85% might be enough)

• Task Assignment Mechanism, Can we

− Select a set of participants before the MCS process, and

− Allow selected participant performing PCS task and returning

results autonomously

• Individual Incentives:

− Using Pay per Participant settings?

22 Jan 201533

Institut Mines-Télécom

CrowdRecruiter Research Outline

■ Research Overview

■ Research Problems

■ Technical Challenges

■ Framework and Algorithms

■ Evaluation and Summary

22 Jan 201534

Institut Mines-Télécom

CrowdRecruiter Research Overview

Individual Energy IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per Participant

Partial Coverage

Task Assignment Minimize Incentives under Specific MCS Data Quality Constraint

Individual Task Execution

One-way Piggyback Sensing using Calls

Data Collection and Aggregation

Individual Participants’ concerns.. MCS Organizer’ concerns..

The same task assignment objectivesWith different energy-saving, incentive, and data quality assumptions/settings

22 Jan 201535

Institut Mines-Télécom

CrowdRecruiter Motivation and

Assumptions

■ Motivations• Reducing Individual Energy Consumption

− One-way Piggyback Crowdsensing using call opportunities

• Minimizing #selected participants (Offline) in order to:

− Reduce overall incentive payment, while

− Meeting MCS Data quality requirement (partial coverage)

■ Assumptions• Individual Incentive Mechanism

− Pay per participant

• MCS Data Quality Requirement

− Ensuring a predefined percentage of subareas (cell tower) being covered per sensing cycle.

22 Jan 201536

Institut Mines-Télécom

CrowdRecruiter Research Objectives

– Input:

• Given the target region as a set of cell towers,

• Given a series of sensing cycles,

• Given the historical call/mobility traces of all volunteers;

– Problem• Selecting a minimal subset of participants from all volunteers, in

order to:

• Ensure a predefined percentage of cell towers being covered by the selected participants.

22 Jan 201537

Institut Mines-Télécom

CrowdRecruiter Research Challenges

• Call/Mobility Prediction using Historical Call/Mobility

Traces. Estimating the coverage achieved by a given set of

participants

• Lowering the complexity of participant set Search:

− NP-hardness of selecting the best participant set

meeting the probabilistic coverage goal

− Using local search algorithm (e.g., Adaptive Greedy) to

approximate the near-optimal participant set

• Proposing the appropriate participant selection metrics

and stopping criteria.

22 Jan 201538

Institut Mines-Télécom

CrowdRecruiter Framework

Each iteration• Selecting an unselected user having the maximal utility (aka utility

function) when combing with users already selected.

Estimating the covering probability of selected users

Returning if meetingthe coverage goal

22 Jan 201539

Institut Mines-Télécom

CrowdRecruiter Core Algorithms

■ Call/Mobility Prediction

• The probability of user u placing at least one call in cell tower t at sensing

cycle i.

■ Utility Calculation

• Utility is calculated as the expectation of number of cell towers being

covered by users in combined set S⋃{U}, where S is set of participants

already selected and U is an unselected user

■ Covering Probability Calculation

• The probability of a predefined number (i.e., T ) of cell towers

being covered by selected users in cycle i (NP-hardness)

22 Jan 201540

and

Institut Mines-Télécom

Near-Optimality of CrowdRecruiter

■ The Utility function i.e., Utility(S) is an submodular set

function

■ According to Nemhauser et al. 1978, The greedy-based

participant search process could achieve (1-e-1)

approximation of Utility maximization.

■ For example

• Supposing the greedy process runs 10 iterations and selects 10

users, and these 10 users could cover 63 cell towers in expectation.

• The best 10-user combination (through enumeration) can cover no

more than 100 cell towers in maximal.

22 Jan 201541

*Nemhauser et al. "Best algorithms for approximating the maximum of a submodular set function." Mathematics of operations research. 1978

Institut Mines-Télécom

Dataset and Evaluation Setups

■ Baseline Algorithms (leveraging

the same adaptive greedy local

search process)

1. MaxMin

• Selecting the user having maximal

minimum of covering probabilities

among all cycles in each iteration

2. MaxCom

• Selecting the user having maximal

complementary with participants

already selected in each iteration

3. MaxCov

• Selecting the user covering the most

of cell towers in each iteration

22 Jan 201542

Evaluation Regions• Business Region

• 45 cell towers• Residential Region

• 86 cell towers• Merged Region

• 131 Cell towers

Institut Mines-Télécom

Results and Comparisons

22 Jan 201543

CR. : CrowdRecruiter

Institut Mines-Télécom

Results and Comparisons

22 Jan 201544

CR. : CrowdRecruiter

Institut Mines-Télécom

Temporal Coverage of CrowdRecruiter

22 Jan 201545

Institut Mines-Télécom

Open Issues of CrowdRecruiter

• MCS Data Quality

− Considering both the number of subareas being

covered and the number of sensor readings obtained

in each subarea

• Individual Incentive Model

− Considering both the payment to each participant and

the payment to each task assignment

• Task Assignment Objectives

− Maximizing Overall Data Quality under Incentive Budget

Constraint?

22 Jan 201546

Institut Mines-Télécom

CrowdTasker Research Outline

■ Research Overview

■ Research Problems

■ Technical Challenges

■ Framework and Algorithms

■ Evaluation and Summary

22 Jan 201547

Institut Mines-Télécom

CrowdTasker Research Overview

IndividualEnergy

IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per Participant +

Task

Budget Partial Coverage + #sensor readings

Task Assignment Maximize MCS Data Quality under Incentive Budget Constraint

Individual Task Execution

One-way Piggyback Sensing using Calls

Data Collection and Aggregation

Individual Participants’ concerns.. MCS Organizer’ concerns..

Compared to Crowd Recruiter, new MCS data quality metrics, over incentive budget, individual incentive models, and new task assignment objectives

22 Jan 201548

Institut Mines-Télécom

CrowdTasker Motivation

■ Motivations• Reducing Individual Energy Consumption

− One-way Piggyback Crowdsensing using call opportunities

• Selecting a set of users and determining in which cycle each

user should participate in the MCS task, in order to:

− Maximize overall MCS Data Quality, while

− Ensuring the total incentive payment not exceeding the

given budget

22 Jan 201549

Institut Mines-Télécom

CrowdTasker Assumptiom

• Base/Bonus Incentive Mechanism

− Base: Pay per participant (e.g., Ba = $50/user)

− Bonus: Pay per task assignment (e.g., Bo = $ 5/ task)

– E.g., for a user with 3 assigned tasks, 50+3*5= $65

• MCS Coverage Quality Metrics

− Threshold: each subarea is given a threshold of sensor

readings E, e.g.,5 readings

− Saturation: for each subarea, supposing x sensor readings

collected in the task, the quality of this area is min{x,E},

– 4 readings4 quality, 7 readings 5 quality

− Overall Estimation: sum of the coverage quality of each

subarea as a whole.

22 Jan 201550

Institut Mines-Télécom

CrowdTasker Research Challenges

• Lowering the complexity of user-cycle combination

set Search using the local search process

− A user-cycle combination identifying assign a task to the user

in the sensing cycle

• Designing a task allocation process which can

− Approximate the “real cost” of each participant and,

− Search the near-optimal set of user-cycle combinations

according to the estimated coverage quality and cost.

22 Jan 201551

Institut Mines-Télécom

CrowdTasker Framework

In the nth iteration, Using adaptive Greedy• Selecting a set of user-cycle combinations

Xn maximizing utilityn(X) while ensuring the budget cost C(Xn)≤ budget

Return Xn-1, if CQE (Xn) ≤ CQE(Xn-1), otherwise update Utilityn+1 using Xn, go to next iteration

Estimating coverage quality CQE(Xn)

22 Jan 201552

Institut Mines-Télécom

CrowdTasker Core Algorithms

■ Call/Mobility Prediction (same as CrowdRecruier)

■ Coverage Quality Estimation

• The expectation of coverage quality achieved by user-cycle combinations

selected in X, where A(Cu, i), identifies if user u is assigned a task in sensing

cycle i.

■ Utility Calculation

• For the first outer-loop iteration, the utility function is the margin of coverage

quality improved by selecting a new user-cycle combination <v,j>

• For the rest outer-loop iteration (nth, n>1), the Utility function ratio of coverage

improvement versus the cost of adding a new user-cycle combination, where

costn(v,j) is the “modular approximation” of the real cost C(X).

22 Jan 201553

Institut Mines-Télécom

Near Optimality of CrowdTasker

■ The Coverage Quality function i.e., CQE(X) and Overall

Incentive Cost function i.e., C(X) are submodular

■ According to lyer et al. 2013, the proposed nested-loop greedy

search process could achieve (α, 1-e-1) approximation of CQE

maximization under budget constraint i.e.,

• Max CQE(X) s.t. C(X) ≤ Budget

■ For example

• Given the settings of 10 euro for Base, 1 euro for bonus, supposing the tasks

allocated by CrowdTasker can acheve 630 overall coverage with 10000 euros

budget

• Then the optimal solution achieved by brute-force enumeration no more 1000

coverage quality with 10000*(10+1)/10= 11000 euros.

22 Jan 201554

*Iyer, Rishabh K., and Jeff A. Bilmes. "Submodular optimization with submodular cover and submodular knapsack constraints." NIPS. 2013.

Institut Mines-Télécom

Evaluation Dataset and Setups

22 Jan 201555

■ Baseline Algorithms (leveraging the

single-loop adaptive greedy local

search, like CrowdRecruiter)

1. MaxCQE• Selecting the user-cycle combination having the

maximal coverage quality improvement in each

iteration

2. MaxUtils• Selecting the user-cycle combination having

maximal coverage quality improvement/cost ratio

(using “real cost”)

3. MaxEnum• Enumerating all possible cycle combination for each

user, Selecting the cycle combination of an

unselected user having the maximal coverage

quality improvement/cost ratio

Evaluation Regions• Business Region

• 45 cell towers• Residential Region

• 86 cell towers• Merged Region

• 131 Cell towers

Institut Mines-Télécom

Results and Comparisons

22 Jan 201556

Incentives Settings• Bo= 1, Ba = 10, 30, 50 and 70

Budget Settings• B=10000,20000,30000Coverage Quality Threshold• E= 1, 3, and 5

Computation Time Comparison

Coverage Quality under budget constraints

Institut Mines-Télécom

Spatial Distribution of Sensor Readings

22 Jan 201557

Using CrowdTasker

Institut Mines-Télécom

Outline

■ Introduction

• Motivation & Background

• State-of-the Art

■ Technical Contribution

• EEMC

• EMC3

• CrowdRecruiter

• CrowdTasker

■ Conclusion

• Summary

• Future Work

22 Jan 201558

Institut Mines-Télécom

Summary of Thesis

■ Our research

• Studying four optimization problems in Mobile

Crowdsensing, addressing energy, incentives and data

quality issues

• Proposing a unified design framework (4 step approach)

and four optimization algorithms (EEMC, EMC3,

CrowdRecruiter and CrowdTasker), addressing four

different optimization objectives.

• Evaluating proposed framework/algorithms using large-

scale real-world mobility dataset, and verifying

effectiveness of our algorithms

22 Jan 201559

Institut Mines-Télécom

Future Work

■ Data Fusion and Processing

• E.g., inferring the sensor readings of uncovered areas using the

sensor readings obtained. (Compressive Crowdsensing!)

■ Considering Privacy in Mobile Crowdsensing

• Investigating different privacy preserved strategies

22 Jan 201560

Institut Mines-Télécom

List of Publications I (Crowdsensing)■ Haoyi Xiong, Daqing Zhang, Guanling Chen, Leye Wang and Vincent

Gauthier, CrowdTasker: Maximizing Coverage Quality in Piggyback Crowdsensing under

Budget Constraint, In Proc. of 13th IEEE International Conference on Pervasive

Computing and Communications (PerCom'15), accepted, 2015. (AR: 15%)

■ Haoyi Xiong, Daqing Zhang, Leye Wang, Hakima Chaouchi, EMC3: Energy-efficient

Data Transfer in Mobile Crowdsensing under Full Coverage Constraint, IEEE

Transactions on Mobile Computing (TMC), preprinted online, 2014. (IF:2.912)

■ Haoyi Xiong, Daqing Zhang, Leye Wang, J.Paul Gibson and Jie Zhu, EEMC: Enabling

Energy-efficient Mobile Crowd-sensing with Anonymous Participants, ACM Transactions

on Intelligent Systems and Technology (TIST), in press, 2014. (IF: 9.39)

■ Daqing Zhang*, Haoyi Xiong*, Leye Wang and Guanling Chen, CrowdRecruiter:

Selecting Participants for Piggyback Crowdsensing under Probabilistic Coverage

Constraint, In Proc. of the 2014 ACM International Joint Conference on Pervasive

and Ubiquitous Computing (UbiComp'14), Seattle, WA. (*co-primary, AR: 12%)

■ Daqing Zhang, Leye Wang, Haoyi Xiong and Bin Guo. 4W1H in Mobile Crowd

Sensing. IEEE Communications Magazine (ComMag), 2014. (IF: 4.46)

■ Leye Wang, Daqing Zhang and Haoyi Xiong, effSense: Energy-Efficient and Cost-

Effective Data Uploading in Mobile Crowdsensing , PUCAA'13 with Ubicomp'13.

■22 Jan 201561

Institut Mines-Télécom

List of Publications II (Mobility Prediction)

■ Haoyi Xiong, Daqing Zhang, Daqiang Zhang, Vincent Gauthier, Kun Yang and

Monique Becker, MPaaS: Mobility Prediction as a Service in Telecom Cloud,

Information Systems Frontiers (ISF), 2014, Springer. (IF: 0.73)

■ Haoyi Xiong, Daqing Zhang, Daqiang Zhang and Vincent Gauthier, Predicting Mobile

Phone User Locations by Exploiting Collective Behavioral Patterns, In Proc. of the 9th

IEEE Conference on Ubiquitous Intelligence and Computing (UIC'12), Fukuoka,

Japan, 2012. (Best Paper Award, AR: 25%)

■ Daqiang Zhang, Daqing Zhang, Haoyi Xiong, Laurence T. Yang and Vincent

Gauthier, NextCell: Predicting Location Using Social Interplay from Cell Phone

Traces, IEEE Transactions on Computers (TC), preprinted, 2014. (IF: 1.473)

■ Daqiang Zhang, Daqing Zhang, Haoyi Xiong, Ching-Hsien Hsu and Athanasios

Vasilakos, BASA: Building Mobile Ad-Hoc Social Networks on Top of Android, IEEE

Network Magazine, 2014. (IF: 3.72)

■ Daqiang Zhang, Min Chen, Mohsen Guizani, Haoyi Xiong and Daqing Zhang, Mobility

Prediction in Telecom Cloud Using Mobile Calls, IEEE Wireless Communication

Magazine, 2014. (IF: 6.524)

22 Jan 201562

Institut Mines-Télécom

Two Involved Projects

22 Jan 201563

EU FP7 SOCIETIES

Excellent project and Finalist of European Tech Cluster Leaders Awards

EU FP7 MONICA

Institut Mines-Télécom

Q&A

■Thanks!

22 Jan 201564

Institut Mines-Télécom

Examples of Mobile/Wearable Sensors

22 Jan 201565

UCSD CITISENS Lapka Sensaris

Institut Mines-Télécom

Case Study (Ne=250, 16h~18h, 14 Dec

2011, Residential Traces)

22 Jan 201566

Institut Mines-Télécom

Near-Optimality of CrowdRecruiter

■ The Utility function of CrowdRecruiter is an submodular

set function

■ The greedy-based participant search process could

achieve (1-e-1) approximation of Utility maximization.

■ For example

• Supposing the greedy process runs 10 iterations and selects 10

users, and these 10 users could cover 63 cell towers in

expectation.

• The best 10-user combination (through enumeration) can cover

no more than 100 cell towers in maximal.

22 Jan 201567

Institut Mines-Télécom

Progress of Participant Selection

Each iteration means a new participant being selected

Fastest Growth Fastest Convergence

22 Jan 2015

Institut Mines-Télécom

Zoom-in the 65—100th iterations

Turning Point

22 Jan 2015

Institut Mines-Télécom

Near Optimality of CrowdTasker

■ The Coverage Quality function and Overall Incentive Cost

function of CrowdTasker are submodular

■ The nested-loop greedy search process could achieve (α,

1-e-1) approximation of coverage quality maximization.

■ For example

• Given the settings of 10 euro for Base, 1 euro for bonus’

• Supposing the tasks allocated by CrowdTasker can acheve 630

overall coverage with 10000 euros budget

• Then the optimal solution achieve by brute-force enumeration

achieve no more 1000 coverage quality with 10000*(10+1)/10=

11000 euros.

22 Jan 201570

Institut Mines-Télécom

EMC3 Motivation and Assumption

■ Motivations• Reducing Individual Energy Consumption

− Two-way Piggyback Crowdsensing using call opportunities

• Minimizing #task assignments in order to:

− Reduce Overall incentive payment, while

− Meeting MCS Data quality requirement (Full Coverage)

■ Assumptions• Individual Incentive Mechanism

− Pay per task assignment

• MCS Data Quality Requirement (beyond EEMC)

− Splitting target region cell towers (subareas)

− Ensuring each cell tower being covered by at least one sensed result every sensing cycle.

22 Jan 201571

Institut Mines-Télécom

EMC3 Contribution Summary

Individual Energy IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per task assignment

#result/cycleFull coverage

Task Assignment Minimize Incentives under Specific MCS Data Quality Constraint

Individual Task Execution

Two-way Piggyback Sensing using Calls

Data Collection and Aggregation

Individual Participants’ concerns.. MCS Organizer’ concerns..

22 Jan 201572

Institut Mines-Télécom

CrowdRecruiter Contribution Summary

Individual Energy IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per Participant

Partial Coverage

Task Assignment Minimize Incentives under Specific MCS Data Quality Constraint

Individual Task Execution

One-way Piggyback Sensing using Calls

Data Collection and Aggregation

Individual Participants’ concerns.. MCS Organizer’ concerns..

22 Jan 201573

Institut Mines-Télécom

CrowdTasker Contribution Summary

IndividualEnergy

IndividualIncentive

OverallIncentive

MCS Data Quality

Task Creation Pay per Participant +

Task

Budget Partial Coverage + #sensor readings

Task Assignment Maximize MCS Data Quality under Incentive Budget Constraint

Individual Task Execution

One-way Piggyback Sensing using Calls

Data Collection and Aggregation

Individual Participants’ concerns.. MCS Organizer’ concerns..

22 Jan 201574

Institut Mines-Télécom

Core Algorithms of EEMC I

■ Next-n-Call Probability Estimation

• Probability of user i placing n calls from time t of cycle k to the

end of cycle k

■ Online Poisson Intensity Estimation

• The Poisson intensity of user i’s calls in cycle k, based on the call

traces up to time t

22 Jan 201575

Institut Mines-Télécom

Core Algorithms of EEMC II

■ Already-Assigned-Fulfiling Probability Estimation

• Probability of N users already assigned tasks but having not yet

returned sensed results (at time t) returning their sensed results

before the end of cycle k (NP-hardness in Calculation)

• Note: when P{Xk,t(Ak-Rk)≥Ne-|Rk|} is lower than a given threshold,

EEMC decides that users already assigned tasks cannot guarantee

to return an expected number (Ne) of results then continues

assigning new tasks.

22 Jan 201576

Institut Mines-Télécom

Core Algorithms of EEMC III

■ Future-Surer-Fulfilling Probability Estimation

• Probability of N users from those, who have already assigned tasks

but having not yet returned sensed results (at time t) or who haven’t

placed any calls yet but have higher probability of placing at least two

calls (i.e., FSui), returning their sensed results before the end of cycle k

• Note: When this probability is higher than a threshold, EEMC decides

that sufficient number of better users will place at least two calls in the

future of cycle k, then drops current user and lefts tasks for future.

22 Jan 201577

Institut Mines-Télécom

Four Steps of MCS Process

• MCS Process [Zhang et al. ’14]

− Step 1. an MCS organizer proposes an MCS task in order to collect

sensed results in the given target region and time-frame

− Step 2. Given the mobile users who are willing to participate in the

MCS tasks, the MCS organizer selects a group of mobile users as

MCS participants.

− Step 3. During the MCS task timeframe, the MCS participants (a)

receive and perform the MCS tasks, and (b) return the sensed

results.

− Step 4. The MCS organizer collects/aggregates sensed results from

large crowds…analyzes…

*Zhang et al. 4W1H of Mobile Crowdsensing, IEEE Communication Magazine

22 Jan 201578