Scan-Statistics-Project-4Y-INSA/Comparaison_of_methods.rmd

---
title: "Comparaison of methods"
output: pdf_document
---

# Scan statistique - Méthode de Monte Carlo et calcul de p-value

## Import libraries
```{r}
library("localScore")
library("latex2exp")
library("Rcpp")
library("caret")
```

## 1. Proposition for simulations under $\mathcal{H}_1$

In this part, we propose a method that simulates a Poisson process under the hypothesis $\mathcal{H}_1$. The idea is to simulate a sample under $\mathcal{H}_0$, and add randomly a subsequence under the alternative hypothesis in this sequence. 
The function `PoissonProcess` creates a sequence of Poisson process of a parameter lambda
```{r}
PoissonProcess <- function(lambda,T) {
  return(sort(runif(rpois(1,lambda*T),0,T)))
}
```

The following function creates a sequence under H0 and add a sequence under H1.
```{r}
SimulationH1 <- function(lambda0, lambda1,T,tau){
    ppH0=PoissonProcess(lambda0,T)
    ppH1.segt=PoissonProcess(lambda1,tau)
    dbt=runif(1,0,T-tau)
    ppH0bis=PoissonProcess(lambda0,T)
    ppH1.repo=dbt+ppH1.segt
    ppH0_avant=ppH0bis[which(ppH0bis<ppH1.repo[1])]
    ppH0_apres=ppH0bis[which(ppH0bis>ppH1.repo[length(ppH1.repo)])]
    ppH1=c(ppH0_avant,ppH1.repo,ppH0_apres)
    return (ppH1)
}
```


`TimeBetweenEvent` compute Time Between Event for a `pp` interval.
```{r}
TimeBetweenEvent <- function(pp){
    n=length(pp)
    tbe=pp[2:n]-pp[1:n1-1]
    tbe=c(0,tbe)
    return (tbe)
}

DataFrame <- function(pp,tbe){
    list=data.frame(ProcessusPoisson=pp, TimeBetweenEvent=tbe)
}
```

## 2. Simulation of the sequences under $\mathcal{H}_0$ via a Monte Carlo Method
In this part, we will try to simulate, using a Monte Carlo method, a set of $10^5$ independant samples, under the assumption that $\lambda=\lambda_0$, hence, that we are under the null hypothesis $\mathcal{H}_0$. 

The function `ScanStat` compute the scan statistic for a sequence, given some parameters $T$ and $\tau$. This function returns the value of the scan stat, and the index of the sequence where it happens
```{r}
ScanStat <- function(pp, T, tau){
    n=length(pp)
    stop=n-length(which(pp>(T-tau)))
    ScanStat=0
    for (i in (1:stop)) {
        x=which((pp>=pp[i])&(pp<=(pp[i]+tau)))
        scan=length(x)
        if (scan>ScanStat) {ScanStat=scan
        max=i}
  }   
    return (c(max,ScanStat))
}
```

We test the scan statistic method for different values of $\lambda_0$. The method of scan statistic we implemented will allow us to have access to the scan test statistic and where it happens in the sequence. 
This function `EmpDistrib` compute the empirical distribution using a Monte Carlo estimator for the scan statistic method. It returns a Data Frame, containing the value of the scan, its probability and the value of its cumulative distribution function. 
```{r}
EmpDistrib <- function(lambda, n_sample,T,tau){
    pp=PoissonProcess(lambda,T)
    scan=c(ScanStat(pp,T, tau)[2])
    index=c(ScanStat(pp,T, tau)[1])
    for (i in 2:(n_sample)){
        pp=PoissonProcess(lambda,T)
        scan=rbind(scan,ScanStat(pp,T, tau)[2])
        index=rbind(index,ScanStat(pp,T, tau)[1])
    }
    min_scan=min(scan)-1
    max_scan=max(scan)
    table1=table(factor(scan, levels = min_scan:max_scan))
    EmpDis=data.frame(cdf=cumsum(table1)/sum(table1), proba=table1/sum(table1), index_scan=min_scan:max_scan)
    EmpDis<-EmpDis[,-2]
    return(EmpDis)
    }
```

This function plot the cumulative distribution function associated to an empirical distribution function
```{r}
Plot_CDF <- function(lambda,n_sample,T,tau){
    Emp=EmpDistrib(lambda,n_sample,T,tau)
    title=TeX(paste(r'(Cumulative distribution function for $\lambda=$)', lambda))
    plot(Emp$index_scan, Emp$cdf,type="s",xlab="Number of occurrences",ylab="Probability", main=title, col="red")
    return(Emp)
}
```
### 2.1. Test of $\mathcal{H}_0: \lambda=\lambda_0$ against $\mathcal{H}_0: \lambda=\lambda_1$, where $\lambda_1 > \lambda_0$ 
In this part, we will test different values for $\lambda_0$ and $\lambda_1$, and compute the probability of occurrence of a certain scan statistic.

```{r}
#Empiricial distribution under H0
n_sample=10**4
lambda0=3
T=10
tau=1
ppH0=PoissonProcess(lambda0,T)
CDF=Plot_CDF(lambda0,n_sample,T,tau)
```

Compute $p$-value for scan statistic of `ppH1` with `Emp`:
```{r}
PValue <- function(Emp,ppH, T, tau){
    scanH1=ScanStat(ppH,T,tau)[2]
    index_scanH1=ScanStat(ppH,T,tau)[1]
    index=Emp$index_scan
    n=length(index)
    if (scanH1< min(Emp$index_scan)){
        return (c(scanH1,1,index_scanH1))
        } else{
            if(min(Emp$index_scan)<scanH1 && scanH1<=max(Emp$index_scan)){
                return(c(scanH1,1-Emp$cdf[scanH1-min(Emp$index_scan)+1],index_scanH1))
            } else{return (c(scanH1,0,index_scanH1))}}
}
```

### 2.2. Simulation under $\mathcal{H}_0$ and computation of p-values
On simule des séquences sous $\mathcal{H}_0$, que l'on stocke. On calcule la valeur de la scan stat et de la p-value, que l'on stocke aussi. On a une séquence de p-valeur des scans et une séquence de score local.
```{r}
NbSeqH0=10000
NbSeqH1=NbSeqH0
DataH0=vector("list")
DataH1=vector("list")
lambda0=2
lambda1=5
T=10
tau=1

#Creation of a sequence that contains the sequence simulated under the null hypothesis
for (i in 1:NbSeqH0) {
    ppi=PoissonProcess(lambda0,T)
    DataH0[[i]]=ppi
}

#Creation of a sequence that contains the sequence simulated under the alternative hypothesis
seqH1begin=c()
for (i in 1:NbSeqH1) {
    pphi=SimulationH1(lambda0, lambda1,T,tau)
    DataH1[[i]]=pphi
}

#Computation of the time between events
TimeBetweenEventList <- function(list,n_list){
    TBE=vector("list",length=n_list)
    for (i in (1:n_list)) {
        ppi=list[[i]]
        ni=length(ppi)
        tbei=ppi[2:ni]-ppi[1:ni-1]
        TBE[[i]]=tbei
    }
    return (TBE)
}
tbe0=TimeBetweenEventList(DataH0,NbSeqH0)
```
We compute the p-value associated to all 10000 sequences, and stock them in a vector. 

```{r}
#We start by computing the empirical distribution for lambda0
Emp = EmpDistrib(lambda0,n_sample,T,tau)
scan = c()
pvalue = c()
index_scan = c()

#Then, we stock the p-value and the 
for (i in 1:NbSeqH0){
    ppi = DataH0[[i]]
    result = PValue(Emp,ppi,T,tau)
    scan = c(scan,result[1])
    pvalue = c(pvalue,result[2])
    index_scan = c(index_scan,result[3])
}

ScS_H0=data.frame(num=(1:NbSeqH0), scan_stat=scan, pvalue_scan=pvalue,class=c(pvalue<0.05)*1) 
ScS_H0
sum(ScS_H0$class[which(ScS_H0$class=='1')])/NbSeqH0
```

```{r}
#We start by computing the empirical distribution for lambda0
scan=c()
pvalue=c()
index_scan=c()

#Then, we stock the p-value and the 
for (i in 1:NbSeqH1){
    ppi=DataH1[[i]]
    result=PValue(Emp,DataH1[[i]],T,tau)
    scan=c(scan,result[1])
    pvalue=c(pvalue,result[2])
    index_scan=c(index_scan,result[3])
}
ScS_H1 = data.frame(num=1:NbSeqH1, scan_stat=scan, pvalue_scan=pvalue, class=(pvalue<0.05)*1, begin_scan=index_scan)
ScS_H1
sum(ScS_H1$class[which(ScS_H1$class=='1')])/NbSeqH1

```

`ScanStatMC` compute local score for `Emp`:
```{r}
ScanStatMC <- function(NbSeq, T, tau, Emp, pp0){
    scan=c()
    pvalue=c()
    index_scan=c()

    for (i in 1:NbSeq){
        ppi=pp0[[i]]
        result=PValue(Emp,ppi,T,tau)
        scan=c(scan,result[1])
        pvalue=c(pvalue,result[2])
        index_scan=c(index_scan,result[3])
    }

    ScS_H0=data.frame(num=(1:NbSeq), scan_stat=scan, pvalue_scan=pvalue,class=c(pvalue<0.05))
    return(ScS_H0)
}
```

## 3. Local score
### 3.1. Distribution of scores via Monte Carlo
`ComputeE` compute `E` coefficient:
```{r}
ComputeE <- function(lambda0, lambda1){
    E = 1
    maxXk = floor(E*(log(lambda1/lambda0)))
    while (maxXk < 3) {
        E = E+1
        maxXk = floor(E*(log(lambda1/lambda0)))
    }

    return (E)
}
```

`ScoreDistribEmpiric` compute score for empiric distribution:
```{r}
ScoreDistribEmpiric <- function(lambda0, lambda1, n_sample, T){
    E = ComputeE(lambda0, lambda1)
    Score = c()
    
    for (i in 1:n_sample){
        ppH0 = PoissonProcess(lambda0,T)
        n1 = length(ppH0)
        tbe0 = ppH0[2:n1]-ppH0[1:n1-1]
        X = floor(E*(log(lambda1/lambda0)+(lambda0-lambda1)*tbe0))
        Score=c(Score,X)
    }
    min_X = min(Score)
    max_X = max(Score)

    P_X = table(factor(Score, levels = min_X:max_X))/sum(table(Score))
    df = data.frame("Score_X" = min(Score):max(Score), "P_X" = P_X)
    df <- df[,-2]

    return (df)
}
```

```{r}

lambda0=5
lambda1=7
distrib_mc=ScoreDistribEmpiric(lambda0,lambda1,10000,T)
score_moyen=mean(distrib_mc[,1])
print(score_moyen)
score_max=max(distrib_mc[,1])
print(score_max)
score_min=min(distrib_mc[,1])
print(score_min)
amplitude=abs(score_max-score_min)
print(amplitude)
E=ComputeE(lambda0, lambda1)
print(E)
barplot(distrib_mc[,2])
```


```{r}
ScoreDistribTheo <- function(lambda0, lambda1, T){
    E = ComputeE(lambda0, lambda1)

    score_max = floor(E*log(lambda1/lambda0))
    

    ## score_min compute
    score_min_c = floor(E*log(lambda1/lambda0)+E*(lambda0-lambda1)*T)
    

    l = seq(score_min_c,score_max,1)
    borne_inf = (l-E*log(lambda1/lambda0))/(E*(lambda0-lambda1))
    borne_sup = (l+1-E*log(lambda1/lambda0))/(E*(lambda0-lambda1))
    proba.l = pexp(rate=lambda0,borne_inf)-pexp(rate=lambda0,borne_sup)
    S = sum(proba.l)
    new.proba.s = proba.l/S
    df = data.frame("Score_X" = l, "P_X" = new.proba.s)

    return (df)
}
```

```{r}

T=10
distrib_score_mc=ScoreDistribEmpiric(2,3,10000,T)

distrib_score_theo=ScoreDistribElisa(2,3,T)

distrib_score_mc
distrib_score_theo


distrib_score_mc = ScoreDistribEmpiric(2,3,10000,T)
distrib_score_theo = ScoreDistribTheo(2,3,T)

plot_graph_distrib_score <- function(distrib_score_theo, distrib_score_mc){
    # length(distrib_score_mc[,2])
    # length(distrib_score_theo[,2])

    #diff_distrib_score=abs(distrib_score_mc[,2]-distrib_score_theo[,2])

    
    #par(mfrow = c(1,2))
    barplot(distrib_score_mc[,2],col="blue",axes=F)
    mtext("Distribution of scores via Monte Carlo",side=1,line=2.5,col="blue")
    axis(2, ylim=c(0,10))
    par(new = T)
    barplot(distrib_score_theo[,2],col="red",axes=F)
    mtext("Distribution of scores using the theoretical method",side=1,line=4,col="red") 
}


plot_graph_distrib_score(distrib_score_theo, distrib_score_mc)
```

### 3.2. Local score calculation
```{r}
LocalScoreMC <- function(lambda0, lambda1, NbSeq, T, X_seq, P_X, tbe0){
  E = ComputeE(lambda0, lambda1)
  
  pvalue = c()
  X = c()
  
  min_X = min(X_seq)
  max_X = max(X_seq)
  
  for (i in 1:NbSeq){
      x = floor(E*log(dexp(tbe0[[i]], rate = lambda1)/dexp(tbe0[[i]], rate = lambda0)))
      X = c(X,x)
      LS = localScoreC(x)$localScore[1]
      
      daudin_result = daudin(localScore = LS, score_probabilities = P_X, sequence_length = length(x), sequence_min = min_X, sequence_max = max_X)
      options(warn = -1) # Disable warnings print
      
      pvalue = c(pvalue, daudin_result)
  }
  LS_H0=data.frame(num=1:NbSeq, pvalue_scan=pvalue, class=(pvalue<0.05))
  return(LS_H0)
}
```

## 4. Experience plan for comparaison
```{r}
NbSeq = 10**2
T = 10
for (lambda0 in (2)){
  Sensitivity = c()
  Specificity = c()
  accepted_lambda = c()

  for (lambda1 in c(3)){
    if (lambda0 < lambda1){
      
      accepted_lambda = c(accepted_lambda,lambda1)
      cat("For T = ", T, ", Nb = ", NbSeq, ", lambda0 = ", lambda0, " and lambda1 = ", lambda1, ":\n", sep = "")
      tbe0 = vector("list",length=NbSeq)
      pp0 =  vector("list", length = NbSeq)
      pp1 =  vector("list", length = NbSeq)
      tbe1 = vector("list", length =  NbSeq)
      
      theoretical_results = c(rep(0,NbSeq), rep(1,NbSeq))
      
      
      for (i in (1:NbSeq)) {
        #Simulation for sequences under H0
        ppi = PoissonProcess(lambda0,T)
        ni=length(ppi)
        pp0[[i]] = ppi
        tbei = ppi[2:ni]-ppi[1:ni-1]
        tbe0[[i]] = tbei
        
        #Simulation for sequences under H1
        ppj1 = SimulationH1(lambda0, lambda1, T, tau)
        nj = length(ppj1)
        pp1[[i]] = ppj1
        tbej = ppj1[2:nj]-ppj1[1:nj-1]
        tbe1[[i]] = tbej
      }
            
      #cat("- Empiric version:\n")
      Score = ScoreDistribEmpiric(lambda0, lambda1, NbSeq, T)
      Emp = EmpDistrib(lambda0,n_sample,T,tau)
      
      X_seq = Score$Score_X
      P_X = Score$P_X
      
      LS_H0 = LocalScoreMC(lambda0, lambda1, NbSeq, T, X_seq, P_X, tbe0)
      options(warn = -1) # Disable warnings print
      SS_H0 = ScanStatMC(NbSeq, T, tau, Emp, pp0)
      SS_H1 = ScanStatMC(NbSeq, T, tau, Emp, pp1)
      
      SS_expected = c(SS_H0$class, SS_H1$class)
            
      #cat("Local Score:\n")
      #print(summary(LS_H0))
      #cat("Scan Statistics:\n")
      #print(summary(SS_H0))
      #cat("Confusion Matrix:\n")
      #print(confusionMatrix(factor(LS_H0$class), factor(SS_H0$class)))
        
      #cat("- Elisa version:\n")
      Score = ScoreDistribTheo(lambda0, lambda1, T)
      Emp = EmpDistrib(lambda0,n_sample,T,tau)

      X_seq = Score$Score_X
      P_X = Score$P_X
            
      LS_H0 = LocalScoreMC(lambda0, lambda1, NbSeq, T, X_seq, P_X, tbe0)
      options(warn = -1) # Disable warnings print

      SS_H0 = ScanStatMC(NbSeq, T, tau, Emp, pp0)
        
      #cat("Local Score:\n")
      #print(summary(LS_H0))
      #cat("Scan Statistics:\n")
      #print(summary(SS_H0))
      #cat("Confusion Matrix:\n")
      print(confusionMatrix(factor(theoretical_results), factor(SS_expected)))
      #Sensitivity = c(Sensitivity,confusionMatrix(factor(theoretical_results), factor(SS_expected))$byClass[1])
      #Specificity = c(Specificity,confusionMatrix(factor(theoretical_results), factor(SS_expected))$byClass[2])

      cat("---\n")
      
    }
  }
  titleSens=TeX(paste(r'(Sensitivity for $\lambda_0=$)', lambda0))
  plot(x=accepted_lambda,y=Sensitivity, type='l', main = titleSens)
  
  titleSpec=TeX(paste(r'(Specificity for $\lambda_0=$)', lambda0))
  plot(x=accepted_lambda,y=Specificity, type='l', main = titleSpec)

}
```

```{r}
theo = c(0,0,0,1,1,1)
exp = c(0,1,1,1,1,0)

confusionMatrix(factor(exp), factor(theo), positive = '1') #prédiction puis théorique
```

```{r}
NbSeq = 10**2
T = 10
lambda0 = 2
lambda1 = 5
n_sample=10**4

cat("For T = ", T, ", Nb = ", NbSeq, ", lambda0 = ", lambda0, " and lambda1 = ", lambda1, ":\n", sep = "")
tbe0 = vector("list",length=NbSeq)
pp0 =  vector("list", length = NbSeq)
pp1 =  vector("list", length = NbSeq)
tbe1 = vector("list", length =  NbSeq)
      
theoretical_results = c(rep(0,NbSeq), rep(1,NbSeq))
      
      
for (i in (1:NbSeq)) {
  #Simulation for sequences under H0
  ppi = PoissonProcess(lambda0,T)
  ni=length(ppi)
  pp0[[i]] = ppi
  tbei = ppi[2:ni]-ppi[1:ni-1]
  tbe0[[i]] = tbei
        
  #Simulation for sequences under H1
  ppj1 = SimulationH1(lambda0, lambda1, T, tau)
  nj = length(ppj1)
  pp1[[i]] = ppj1
  tbej = ppj1[2:nj]-ppj1[1:nj-1]
  tbe1[[i]] = tbej
  }
            
      Emp = EmpDistrib(lambda0,n_sample,T,tau)

      SS_H0 = ScanStatMC(NbSeq, T, tau, Emp, pp0)
      SS_H1 = ScanStatMC(NbSeq, T, tau, Emp, pp1)
      
      SS_expected = c(SS_H0$class, SS_H1$class)
            
      #cat("Local Score:\n")
      #print(summary(LS_H0))
      #cat("Scan Statistics:\n")
      #print(summary(SS_H0))
      #cat("Confusion Matrix:\n")
      #print(confusionMatrix(factor(LS_H0$class), factor(SS_H0$class)))
        
      #cat("- Elisa version:\n")
      Score = ScoreDistribTheo(lambda0, lambda1, T)
      Emp = EmpDistrib(lambda0,n_sample,T,tau)

      X_seq = Score$Score_X
      P_X = Score$P_X
            
      LS_H0 = LocalScoreMC(lambda0, lambda1, NbSeq, T, X_seq, P_X, tbe0)
      options(warn = -1) # Disable warnings print

      SS_H0 = ScanStatMC(NbSeq, T, tau, Emp, pp0)
        
      #cat("Local Score:\n")
      #print(summary(LS_H0))
      #cat("Scan Statistics:\n")
      #print(summary(SS_H0))
      #cat("Confusion Matrix:\n")
      print(confusionMatrix(factor(theoretical_results), factor(SS_expected)))
      #Sensitivity = c(Sensitivity,confusionMatrix(factor(theoretical_results), factor(SS_expected))$byClass[1])
      #Specificity = c(Specificity,confusionMatrix(factor(theoretical_results), factor(SS_expected))$byClass[2])

      cat("---\n")
      
  titleSens=TeX(paste(r'(Sensitivity for $\lambda_0=$)', lambda0))
  plot(x=accepted_lambda,y=Sensitivity, type='l', main = titleSens)
  
  titleSpec=TeX(paste(r'(Specificity for $\lambda_0=$)', lambda0))
  plot(x=accepted_lambda,y=Specificity, type='l', main = titleSpec)
```