Got predict-victims running in parallel, drastically reducing the time

for each test. Also changed how we get the rankings of infected
individuals each day.

4 files changed, 60 insertions, 43 deletions

diff --git a/R Scripts/find-parents.R b/R Scripts/find-parents.R
index 3ec8809..023d7ba 100644
--- a/R Scripts/find-parents.R
+++ b/R Scripts/find-parents.R
@@ -6,8 +6,8 @@
 # source('criminal_cascades/R Scripts/structural.R')
 
 ##### Initialize parameters based on what ml2 found
-alpha = 0.061
-delta = 0.082
+alpha = 0.18
+delta = 0.09
 
 ##### Get weights
 edges = dag_dat_test[!is.na(dag_dat_test$t2),]
@@ -21,20 +21,27 @@ weights = p/p_tilde
 edges$weight = weights
 
 ##### Find most likely parents
-parents = data.frame(vic=0,Npars=0,par_rank=0)
+parents = data.frame(vic=0,Npars=0,par_rank=0,rand_rank=0)
 vics = setdiff(vic_ids,seeds)
+print(length(vics))
 for (u in vics){
+  if(which(vics==u) %% 500 == 0) print(which(vics==u))
   u_parents = edges[edges$to==u,]
   u_parents = u_parents[order(u_parents$weight,decreasing=T),]
   Nparents = dim(u_parents)[1]
   infector = V(g)$infector[u]
   infectorID = which(u_parents$from==infector)
-  parents[which(vics==u),] = c(u, Nparents, infectorID)
+  randID = sample(1:Nparents,1)
+  parents[which(vics==u),] = c(u, Nparents, infectorID, randID)
 }
 
 ##### Get some summary statistics on how well 
+mean(parents$par_rank==1)
 median(parents$par_rank[parents$Npars>9])
 median(parents$par_rank[parents$Npars>99])
 
+edges[edges$to==2847,]
 
+## baseline alg
+# for each vic, find potential parents, pick one at random
 
diff --git a/R Scripts/generate-network.R b/R Scripts/generate-network.R
index 3b40969..dab81a4 100644
--- a/R Scripts/generate-network.R
+++ b/R Scripts/generate-network.R
@@ -3,20 +3,20 @@ setwd("~/Documents/Cascade Project/")
 source('criminal_cascades/R Scripts/temporal.R')
 source('criminal_cascades/R Scripts/structural.R')
 
-alpha = 1/100
-beta = 0.02
-delta = 0.15
+alpha = 1/10
+beta = 0.01
+delta = 0.25
 # lmbda = 1/10
 t_max = 1000
 
 N = 5000
 g = forest.fire.game(nodes=N, fw.prob=0.3, ambs=1, directed=F)
-plot(g, vertex.size=5, vertex.label=NA)
+plot(g, vertex.size=3, vertex.label=NA)
 
 V(g)$seed = runif(vcount(g))<beta
 seeds = which(V(g)$seed)
 V(g)$vic = V(g)$seed
-V(g)$vic.day[V(g)$seed] = sample(1:t_max, sum(V(g)$seed))
+V(g)$vic.day[V(g)$seed] = 1#sample(1:t_max, sum(V(g)$seed))
 V(g)$spawn.date = 0
 V(g)$infector = NA
 
@@ -28,17 +28,19 @@ for (day in 1:t_max){
     dists = as.numeric(shortest.paths(g,vic,neighbors))
     infected = neighbors[which(runif(length(neighbors))<structural(delta, dists))]
     infected = setdiff(infected,seeds) # don't try to infect seeds
-    inf.days = day + ceiling(alpha*rexp(length(infected),alpha))
+    inf.days = day + ceiling(rexp(length(infected),alpha))
+    realized = ((inf.days <= V(g)$vic.day[infected]) %in% c(NA,T)) & (inf.days<=t_max)
+    infected = infected[realized]
     V(g)$vic[infected] = TRUE
-    infects = (inf.days <= V(g)$vic.day[infected]) %in% c(NA,T)
-    V(g)$vic.day[infected[infects]] = inf.days[infects]
-    V(g)$infector[infected[infects]] = vic
+    V(g)$vic.day[infected] = inf.days[realized]
+    V(g)$infector[infected] = vic
   }
 }
 
 vic_ids = which(V(g)$vic)
+print(length(vic_ids))
 cols = rep('lightblue',N); cols[V(g)$vic]='red'; cols[V(g)$seed]='darkred'
-plot(g, vertex.size=5, vertex.label=NA, vertex.color=cols)
+plot(g, vertex.size=3, vertex.label=NA, vertex.color=cols)
 
 ##### generate dag_dat
 dag_dat_test = data.frame(matrix(nrow=1,ncol=10))
@@ -68,12 +70,12 @@ rownames(dag_dat_test) = NULL
 
 write.csv(dag_dat_test, file='Results/dag_dat_test.csv')
 
-##### analyze performance of recovery algorithm
-recovered = read.csv('Results/infectors.csv',header=F,col.names=c('victim','infector'))
-recovered = recovered[order(recovered$victim),]
-infectors = cbind(setdiff(vic_ids,seeds),
-                  V(g)$infector[setdiff(vic_ids,seeds)],
-                  recovered$infector[recovered$victim %in% setdiff(vic_ids,seeds)])
-mean(infectors[,2]==infectors[,3])
-
-dag_dat_test[dag_dat_test$to==4984,]
\ No newline at end of file
+##### analyze performance of recovery algorithm ------
+# recovered = read.csv('Results/infectors.csv',header=F,col.names=c('victim','infector'))
+# recovered = recovered[order(recovered$victim),]
+# infectors = cbind(setdiff(vic_ids,seeds),
+#                   V(g)$infector[setdiff(vic_ids,seeds)],
+#                   recovered$infector[recovered$victim %in% setdiff(vic_ids,seeds)])
+# mean(infectors[,2]==infectors[,3])
+# 
+# dag_dat_test[dag_dat_test$to==4984,]
diff --git a/R Scripts/predict-victims-plots.R b/R Scripts/predict-victims-plots.R
index 8a93667..2ac62c8 100644
--- a/R Scripts/predict-victims-plots.R
+++ b/R Scripts/predict-victims-plots.R
@@ -2,7 +2,7 @@
 hist(correct_rank3,150,xlim=c(0,vcount(lcc)),col=rgb(0,0,1,1/8),
      xlab='Risk Ranking of Victims',main='')
 hist(correct_rank1,150,xlim=c(0,vcount(lcc)),col=rgb(1,0,1,1/8),add=T)
-hist(correct_rank2,150,xlim=c(0,vcount(lcc)),col=rgb(1,0,1,1/8),add=T)
+hist(correct_rank2,150,xlim=c(0,vcount(lcc)),col=rgb(0,0,1,1/8),add=T)
 legend("topright", c("Demographics Model", "Cascade Model"), 
        fill=c(rgb(1,0,1,1/8), rgb(0,0,1,1/8)))
 
@@ -12,9 +12,9 @@ counts = matrix(c(colSums(correct_rank<(vcount(lcc)/1000))*100/nvics,
                 nrow=3, byrow=T)
 plot(lambdas,counts[1,],log='x',type='l')
 
-correct_rank1 = correct_rank[,length(lambdas)]
-correct_rank2 = correct_rank[,1]
-correct_rank3 = correct_rank[,which.min(colMeans(correct_rank))]
+correct_rank1 = correct_rank[,length(lambdas)] # demographics model
+correct_rank2 = correct_rank[,1] # cascade model
+correct_rank3 = correct_rank[,which.min(colMeans(correct_rank))] # best combined model
 counts = matrix(c(sum(correct_rank1<(vcount(lcc)*0.001)),
                   sum(correct_rank1<(vcount(lcc)*0.005)),
                   sum(correct_rank1<(vcount(lcc)*0.01)),
@@ -59,3 +59,4 @@ legend("bottomright", inset=0.05,
        c("Demographics Model", "Cascade Model", "Combined Model"), 
        fill=c('red','darkblue','darkgreen'))
 lines(c(0,vcount(lcc)),c(0,1))
+
diff --git a/R Scripts/predict-victims.R b/R Scripts/predict-victims.R
index 470815d..2bda7e2 100644
--- a/R Scripts/predict-victims.R
+++ b/R Scripts/predict-victims.R
@@ -1,4 +1,7 @@
 library(igraph)
+library(foreach)
+library(doMC)
+registerDoMC(cores=4)
 setwd('~/Documents/Cascade Project')
 load('Raw Data/lcc.RData')
 load('Results/hyper-lcc.RData')
@@ -7,34 +10,36 @@ source('criminal_cascades/R Scripts/temporal.R')
 source('criminal_cascades/R Scripts/structural.R')
 
 ##### Initialize data
-formula = vic ~ sex + race + age + gang.member + gang.name
+formula = vic ~ sex + race + age + gang.member #+ gang.name
 lcc_verts$sex = as.factor(lcc_verts$sex)
 lcc_verts$race = as.factor(lcc_verts$race)
 lcc_verts$age = as.numeric(lcc_verts$age)
 lcc_verts$gang.name = as.factor(lcc_verts$gang.name)
 # sum(hyp_lcc_verts$vic)/length(days)
+df = data.frame(ir=lcc_verts$ir_no, dem=0, cas=0, comb=0)
 
 alpha = 0.0028
 delta = 0.06
 days = sort(unique(hyp_lcc_verts$vic.day)) # 70:max(hyp_lcc_verts$vic.day, na.rm=T)
 lambdas = c(0,1)#c(0, exp(seq(log(0.0000001), log(.0005), length.out=150)), 1)
-nvics = sum(lcc_verts$vic)#sum(hyp_lcc_verts$vic.day %in% days)
-correct_rank = matrix(nrow=nvics, ncol=length(lambdas))
+nvics = sum(hyp_lcc_verts$vic.day %in% days)
 edges_all = dag_dat_all
 
 ##### Loop through days
+writeLines(c(""), "Results/log.txt")
 ptm = proc.time()
-for (day in days){
-  if (which(day==days) %% 100 == 0) print(day)
-  
+correct_rank = foreach (day = days, .combine=rbind) %dopar% {
+  if (which(day==days) %% 100 == 0){sink("Results/log.txt", append=TRUE);cat(paste("day:",day,"\n"))}
+
   ##### Demographics model
   vics = match(unique(hyp_lcc_verts$ir_no[which(hyp_lcc_verts$vic.day<day)]),lcc_verts$name)
   victims = lcc_verts[,c('vic','sex','race','age','gang.member','gang.name')]
   victims$vic[vics] = TRUE
   victims$vic[-vics] = FALSE
-#   glm.fit = glm(formula, data=victims, family=binomial)
-  glm.fit = lm(formula, data=victims)
-  glm.probs = predict(glm.fit, newdata=lcc_verts, type='response')
+  fit = lm(formula, data=victims)
+#   fit = glm(formula, data=victims, family=binomial)
+#   fit = randomForest(formula, data=victims[,1:5], ntree=100)
+  probs = predict(fit, newdata=lcc_verts, type='response')
   
   ##### Cascade Model
   edges = edges_all[which(edges_all$t1<day),]
@@ -49,19 +54,21 @@ for (day in days){
 # maybe need to change this to reflect new algorithm that accounts for \tilde{p}
 
   ##### Combined Model
-  combined = data.frame(ir=attr(glm.probs,'name'), dem=as.numeric(glm.probs), cas=0, comb=0)
-  combined$cas[match(risk$ir, attr(glm.probs,'name'))] = risk$weight
+  combined = df#data.frame(ir=attr(probs,'name'), dem=as.numeric(probs), cas=0, comb=0)
+  combined$dem[match(attr(probs,'name'), df$ir)] = as.numeric(probs)
+  combined$cas[match(risk$ir, attr(probs,'name'))] = risk$weight
   
   ##### Gather results
   infected_irs = hyp_lcc_verts$ir_no[which(hyp_lcc_verts$vic.day==day)]
+  crday = matrix(nrow=length(infected_irs), ncol=length(lambdas))
   for (lambda in lambdas){
     combined$comb = lambda*combined$dem + (1-lambda)*combined$cas
     c_idx = which(lambdas==lambda)
-    r_idx = head(which(is.na(correct_rank[,c_idx])),length(infected_irs))
-    # !! order should be first: rank of (3,5,5,7) should be (1,2,2,4), may need to do n-rank
-    correct_rank[r_idx,c_idx] = match(infected_irs, combined$ir[order(combined$comb, decreasing=T)])
-    # maybe should also mark down vic/nonvic status of each?
+    crday[,c_idx] = rank(-combined$comb,ties.method='average')[match(infected_irs,combined$ir)]
   }
-  
+
+  return(crday)
 }
 print(proc.time()-ptm)
+
+# save(correct_rank, file='Results/correct_rank_62815.RData') 
\ No newline at end of file