example3: Example of using 'deduplication' package - the case of...

In bogdanoancea/deduplication: R package for computing the duplicity probility for each mobile device.

Description

Example of using deduplication package - showing the case of incompatibility between the type of connection in the network and the emission model.

Usage

example3()

Details

This is an example of how to force the use of an emission model that differs from the type of connection done by the network. Thanks to the information from the simulator and the flexibility of the functions of this package, it is possible to make the assumption of a kind of emission model different from the real type of connection. Obviously, this incoherence has consequences. It could cause to type of problems. Firstly there are events to antennas without signal under the assumption made to build the emission probability matrix. Secondly, there are transitions between pairs of antennas which are not possible in consecutive times because there is some tile without signal between those antennas under the assumption made to build the emission probability matrix. These two problems can be detected by two functions in this package: checkConnections_step1() and checkConnections_step2(). In the following example, one of these kinds of cases is shown.

References

https://github.com/MobilePhoneESSnetBigData

Examples

###    READ DATA    ####

# Set the folder where the necessary input files are stored

path_root <- 'extdata'

 # 0. Read simulation params

simParams <- readSimulationParams(system.file(path_root, 'simulation.xml', 
package = 'deduplication'))

 # 1. Read grid parameters

gridParams <- readGridParams(system.file(path_root, 'grid.csv', 
package = 'deduplication'))

 # 2. Read network events

events <- readEvents(system.file(path_root, 'AntennaInfo_MNO_MNO1.csv', 
package = 'deduplication'))

# 3. Set the signal file name, i.e. the file where the signal strength/quality 
# for each tile in the grid is stored

signalFileName <- system.file(path_root, 'SignalMeasure_MNO1.csv')

 # 4. Set the antennas file name

antennasFileName <- system.file(path_root, 'antennas.xml')

####    PREPARE DATA    ####

 # 5. Initial state distribution (prior)

nTiles <- gridParams$ncol * gridParams$nrow 
initialDistr_RSS_uniform.vec <- initialDistr_SDM_uniform.vec <- rep(1 /
nTiles, nTiles)

 # 6. Get a list of detected devices

 devices <- getDeviceIDs(events)

 # 7. Get connections for each device

connections <- getConnections(events)

# 8. Emission probabilities are computed from the signal strength/quality file.  
# In this case handoverType is strength then the emission model should be RSS.

emissionProbs_RSS <- getEmissionProbs(nrows = gridParams$nrow, 
ncols = gridParams$ncol, signalFileName = signalFileName, 
sigMin = simParams$conn_threshold, handoverType = simParams$connection_type[[1]], 
emissionModel = "RSS", antennaFileName = antennasFileName)

 # We also try to use the emission model with the SDM assumption.

emissionProbs_SDM <- getEmissionProbs(nrows = gridParams$nrow, 
ncols = gridParams$ncol, signalFileName = signalFileName, 
sigMin = simParams$conn_threshold, handoverType = simParams$connection_type[[1]], 
emissionModel = "SDM", antennaFileName = antennasFileName)

 # 9. Build joint emission probabilities for both options: RSS and SDM.

jointEmissionProbs_RSS <- getEmissionProbsJointModel(emissionProbs_RSS)

jointEmissionProbs_SDM <- getEmissionProbsJointModel(emissionProbs_SDM)

 # 10. Build the generic model for both cases and the a priori 
 # uniform (by default)

model_RSS_uniform <- getGenericModel( nrows = gridParams$nrow, 
ncols = gridParams$ncol, emissionProbs_RSS, initSteady = FALSE, 
aprioriProb = initialDistr_RSS_uniform.vec)

model_SDM_uniform <- getGenericModel( nrows = gridParams$nrow, 
ncols = gridParams$ncol, emissionProbs_SDM, initSteady = FALSE, 
aprioriProb = initialDistr_SDM_uniform.vec)

 # 11. Fit models.

ll_RSS_uniform <- fitModels(length(devices), model_RSS_uniform, connections)

ll_SDM_uniform <- fitModels(length(devices), model_SDM_uniform, connections)

# The log likelihood is infinity for those connections that are impossible 
# under the model SDM (ll_SDM_uniform).


 # 12. Make the checks and corresponding imputations to force the fit of SDM model.

### check1: connections incompatibles with emissionProbs are imputed as NA

check1_SDM <- checkConnections_step1(connections, emissionProbs_SDM)

check1_SDM$infoCheck_step1 connections_ImpSDM0 <-check1_SDM$connectionsImp

### check2: jumps incompatibles with emissionProbs are avoid by making time padding.

check2_SDM <- checkConnections_step2(emissionProbs = emissionProbs_SDM, 
connections = connections_ImpSDM0, gridParams = gridParams)

connections_ImpSDM <- check2_SDM$connections_pad


 # 13. Fit models.

ll_SDM_uniform <- fitModels(length(devices), model_SDM_uniform, connections_ImpSDM)

 # 14. Build the joint model.

modelJ_RSS_uniform <- getJointModel( nrows = gridParams$nrow, 
ncols = gridParams$ncol, jointEmissionProbs = jointEmissionProbs_RSS, 
initSteady = FALSE, aprioriJointProb = initialDistr_RSS_uniform.vec)

modelJ_SDM_uniform <- getJointModel( nrows = gridParams$nrow, 
ncols = gridParams$ncol, jointEmissionProbs = jointEmissionProbs_SDM, 
initSteady = FALSE, aprioriJointProb = initialDistr_SDM_uniform.vec)

 # 15. Apriori probability of 2-to-1

Pii <- aprioriOneDeviceProb(simParams$prob_sec_mobile_phone, length(devices))

 # 16. Build a matrix of pairs of devices to compute duplicity probability

pairs4dup_RSS <- computePairs(connections, length(devices), oneToOne = TRUE)

pairs4dup_SDM <- computePairs(connections_ImpSDM, length(devices), oneToOne = TRUE)


####    COMPUTE DUPLICITY    ####


# 17. Compute duplicity probabilities using "1to1" method

duplicity1to1_RSS_uniform.dt <- computeDuplicityBayesian("1to1", devices, 
pairs4dup_RSS, modelJ_RSS_uniform, ll_RSS_uniform, P1 = NULL, Pii=Pii)

duplicity1to1_SDM_uniform.dt <- computeDuplicityBayesian("1to1", devices,
pairs4dup_SDM, modelJ_SDM_uniform, ll_SDM_uniform, P1 = NULL, Pii=Pii)

bogdanoancea/deduplication documentation built on Dec. 2, 2020, 11:22 p.m.