In this chapter, we will be using the data scraped in the part 1 of the book. The data has text that describes profiles of freelancers, and the hourly rate they are charging in USD. The idea of the following section is to fit a model that given the skills of a freelancer, we are able to predict its hourly salary.

The following code shows how to convert the raw text that in this case has skills of a user in a bag of words matrix. For this we use an R library called tm. This means that for each word in the corpus we create variable with the amount of occurrences of each variable.

library(tm)
library(data.table)  

source(''text_analytics/text_analytics_functions.R'') 
data = fread(''text_analytics/data/profiles.txt'') 
rate = as.numeric(data$rate) 
keep = !is.na(rate) 
rate = rate[keep]  

### Make bag of words of title and body 
X_all = bag_words(data$user_skills[keep]) 
X_all = removeSparseTerms(X_all, 0.999) 
X_all 

# <<DocumentTermMatrix (documents: 389, terms: 1422)>> 
#   Non-/sparse entries: 4057/549101 
# Sparsity           : 99% 
# Maximal term length: 80 
# Weighting          : term frequency - inverse document frequency (normalized) (tf-idf) 

### Make a sparse matrix with all the data 
X_all <- as_sparseMatrix(X_all)

Now that we have the text represented as a sparse matrix we can fit a model that will give a sparse solution. A good alternative for this case is using the LASSO (least absolute shrinkage and selection operator). This is a regression model that is able to select the most relevant features to predict the target.

train_inx = 1:200
X_train = X_all[train_inx, ] 
y_train = rate[train_inx]  
X_test = X_all[-train_inx, ] 
y_test = rate[-train_inx]  

# Train a regression model 
library(glmnet) 
fit <- cv.glmnet(x = X_train, y = y_train,  
   family = ''gaussian'', alpha = 1,  
   nfolds = 3, type.measure = ''mae'') 
plot(fit)  

# Make predictions 
predictions = predict(fit, newx = X_test) 
predictions = as.vector(predictions[,1]) 
head(predictions)  

# 36.23598 36.43046 51.69786 26.06811 35.13185 37.66367 
# We can compute the mean absolute error for the test data 
mean(abs(y_test - predictions)) 
# 15.02175

Now we have a model that given a set of skills is able to predict the hourly salary of a freelancer. If more data is collected, the performance of the model will improve, but the code to implement this pipeline would be the same.