Econ B2000, MA Econometrics

Kevin R Foster, the Colin Powell School at the City College of New York, CUNY

Fall 2019

For this lab, we will use simple k-nn techniques of machine learning to try to guess people’s neighborhoods.

Form a group of 3 (again, invite someone new). Groups should prepare a 4-min presentation by one of the group members about their experiment process and results. You get 45 min to prepare.

You’ve probably done this in your ordinary life: meet someone and guess what neighborhood they live in. Here we try to train the computer, using the PUMS data again.

Start with looking at the differences in means of some of the variables and put that together with your own knowledge of the city. You might want to subset the data – are you trying to predict everbody? Are millenials easier? Retirees? College grads?

Then move toward a k-nn classification. Start by just trying to predict the borough. Create this factor:

dat_NYC <- subset(acs2017_ny, (acs2017_ny$in_NYC == 1)&(acs2017_ny$AGE > 20) & (acs2017_ny$AGE < 66))
attach(dat_NYC)
borough_f <- factor((in_Bronx + 2*in_Manhattan + 3*in_StatenI + 4*in_Brooklyn + 5*in_Queens), levels=c(1,2,3,4,5),labels = c("Bronx","Manhattan","Staten Island","Brooklyn","Queens"))

What variables do we think are relevant in classifying by borough? NOT PUMA since neighborhood likely perfectly classifies… I’ll give an example, where I try income_total, owner_cost combined with rent_cost. You should find other data to do better.

Here is some code to get you started. Best to normalize – here is a bit of code to get the data to have a mean of zero and stdev of one.

norm_varb <- function(X_in) {
(X_in - mean(X_in, na.rm = TRUE))/sd(X_in, na.rm = TRUE)
}

Next, fix up the data,

is.na(OWNCOST) <- which(OWNCOST == 9999999)
housing_cost <- OWNCOST + RENT
norm_inc_tot <- norm_varb(INCTOT)
norm_housing_cost <- norm_varb(housing_cost)

Here we create the dataframe to use,

data_use_prelim <- data.frame(norm_inc_tot,norm_housing_cost)
good_obs_data_use <- complete.cases(data_use_prelim,borough_f)
dat_use <- subset(data_use_prelim,good_obs_data_use)
y_use <- subset(borough_f,good_obs_data_use)

Next split the data into 2 parts: one part to train the algo, then the other part to test how well it works for new data. Here we use an 80/20 split.

set.seed(12345)
NN_obs <- sum(good_obs_data_use == 1)
select1 <- (runif(NN_obs) < 0.8)
train_data <- subset(dat_use,select1)
test_data <- subset(dat_use,(!select1))
cl_data <- y_use[select1]
true_data <- y_use[!select1]

Finally run the k-nn algo and compare against the simple means,

summary(cl_data)
prop.table(summary(cl_data))
summary(train_data)
require(class)
for (indx in seq(1, 9, by= 2)) {
 pred_borough <- knn(train_data, test_data, cl_data, k = indx, l = 0, prob = FALSE, use.all = TRUE)
num_correct_labels <- sum(pred_borough == true_data)
correct_rate <- num_correct_labels/length(true_data)
print(c(indx,correct_rate))
}

Now find some other data that will do a better job of classifying. How good can you get it to be? At what point do you think there might be a tradeoff between better classifying the training data and doing worse at classifying the test data?