Load the Cleansed Dataset

load("SQF_clean.rda")

What to predict?

To begin with, we need to identify a task. Recall from previous sections that cs_* stands for the reasons for stop. Can we generate a model to predict whether a pedestrian will be arrested based on the following feautres?

(You should perform your reasoning for feautre selection. )

Now let’s try out some classification packages in R.

Training and Testing split

Let’s simply divide the dataset into two sets: training and testing sets, where data in testing set will not occur in training set. And we evaluate the performance of our predicitive models.

set.seed(5)
dat$arstmade = as.factor(dat$arstmade)
for(b in grep("cs_", colnames(dat), value=TRUE)) dat[[b]] <- as.factor(dat[[b]])
select <- grep("cs_", colnames(dat), value=TRUE)
select <- c(select, c("sex","race","city","timestop","arstmade"))
dat <- dat[,select]
dat <- dat[complete.cases(dat),]
index = sample(nrow(dat), nrow(dat)*.75)
train = dat[index,]
test = dat[-index,]

In this case, training set has 34340 instances and testing set has 11447.

Metrics: Compute the accuracy, recall, precision, and F-measure

performance <- function(pred, truth){
  cm <- table(pred,truth)      
  tp <- cm[2,2]   
  tn <- cm[1,1]   
  fp <- cm[2,1]   
  fn <- cm[1,2]   
  accuracy <- (tp+tn) / length(truth)   
  recall <- (tp) / (tp+fn)   
  precision <- (tp) / (tp+fp)   
  fmeasure <- 2*recall*precision/(recall+precision)      
  result <- as.data.frame(c(accuracy, recall, precision, fmeasure))
  colnames(result) <- c("Performance")
  row.names(result) <- c("Accuracy", "Recall", "Precision", "F-measure")
  return(result)
}

Decision Tree

library(rpart)

m <- rpart(arstmade ~ sex + race + city + timestop + cs_objcs + cs_descr + cs_casng + cs_lkout + cs_cloth + cs_drgtr + cs_furtv + cs_vcrim + cs_bulge, data=train, control = rpart.control(cp=.0001, minsplit=10))

pred <- predict(m,test,type="class")
truth <- test$arstmade
performance(pred, truth)
##           Performance
## Accuracy    0.8458985
## Recall      0.1192661
## Precision   0.4770642
## Fmeasure    0.1908257

Naive Bayes

library(e1071)

m <- naiveBayes(arstmade ~ sex + race + city + timestop + cs_objcs + cs_descr + cs_casng + cs_lkout + cs_cloth + cs_drgtr + cs_furtv + cs_vcrim + cs_bulge, data=train)

pred <- predict(m,test,type="class")
truth <- test$arstmade
performance(pred, truth)
##           Performance
## Accuracy   0.84747095
## Recall     0.03268349
## Precision  0.49137931
## F-measure  0.06129032

Support Vector Machine

library(doParallel)

registerDoParallel()
getDoParWorkers()
## [1] 4
library(e1071)

m <- svm(arstmade ~ sex + race + city + timestop + cs_objcs + cs_descr + cs_casng + cs_lkout + cs_cloth + cs_drgtr + cs_furtv + cs_vcrim + cs_bulge, data=train, cost=100)

pred <- predict(m,test)
truth <- test$arstmade
performance(pred, truth)
##           Performance
## Accuracy   0.84886870
## Recall     0.03841743
## Precision  0.55833333
## F-measure  0.07188841

Random Forest

library(randomForest)

m <- randomForest(arstmade ~ sex + race + city + timestop + cs_objcs + cs_descr + cs_casng + cs_lkout + cs_cloth + cs_drgtr + cs_furtv + cs_vcrim + cs_bulge, data=train)

pred <- predict(m,test,type="class")
truth <- test$arstmade
performance(pred, truth)
##           Performance
## Accuracy   0.85044116
## Recall     0.04357798
## Precision  0.63333333
## F-measure  0.08154506