I wrote about an ‘algorithm’ that I concocted in my head for determining “Drop the Mic” winners in an earlier post: (“Drop the Mic” winner algorithm!). In that post I found that my silly little algorithm has accurately predicted the winner 85% of the time. 👍

To rehash that algorithm I said:

1. If one contestant is female and the other is male, the female wins!
2. If both are male, or both are female, or if they are a group, the one standing to the left of the screen at the end wins!

That was fun, but it wasn’t very data science-y. So… let’s apply a more proper machine learning method to determine the winner algorithm.

We start again with scraping the data from IMDb:

library(rvest)
library(tidyverse)
library(knitr)
library(kableExtra)

# Scrape data from IMDb
data.dropmic <- read_html("http://www.imdb.com/title/tt6396094/episodes?season=1&ref_=tt_eps_sn_1") %>%
        html_nodes("strong a") %>%
        html_text() %>%
        as.data.frame() %>%
        .[grepl(" vs", .[,1]),] %>%
        data.frame(Episode = .)

data.clean <- str_split(data.dropmic$Episode, "&| and ") %>%
        unlist() %>%
        data.frame(Battles = .)

# Display results
kable(data.clean, "html") %>%
        kable_styling(bootstrap_options = c("striped", "hover"))

The Battles were properly scraped in the above table. Let’s grab the following features from it:

1. Who are the contestants?
2. What are the genders of the contestants (we shall designate the first name in the battle as contestant 1)?
3. Where is contestant 1 standing (left or right on the screen)?

# One data cleaning step needs to be done... make all "vs. " to "vs "
data.clean$Battles <- gsub("vs. ", "vs ", data.clean$Battles)

## 1. Who are the contestants?
# Separate contestants
features <- str_split(data.clean$Battles, "vs ")

# Use purrr::map to add Contestants to data.clean
data.model <- cbind(data.clean, 
                    Contestant.1 = purrr::map(features, 1) %>% unlist(),
                    Contestant.2 = purrr::map(features, 2) %>% unlist())

# Display results
kable(data.model, "html") %>%
        kable_styling(bootstrap_options = c("striped", "hover"))

We have our contestants! Let’s now add the genders of both contestants (I’m doing this manually… training a convolutional neural network to recognize females is a bit out of scope for this post 😛) as well as the positions and note the winners:

Nice! We are now ready to train a machine learning model to predict the winner:

library(caret)
set.seed(1)

# Create training dataset
train.index <- createDataPartition(data.model.comp$winner, 
                                   p = 0.8,
                                   list = F)

data.train <- data.model.comp[train.index,]
data.test <- data.model.comp[-train.index,]

## Establish caret parameters
# Training controls
caret.train.control <- trainControl(
        method = "repeatedcv",
        number = 10,
        repeats = 10
        )
# Training machine learning model
doParallel::registerDoParallel(cores = 6)

machine.model <- train(winner ~ Con1.gender + Con2.gender + Con1.position, 
                       data = data.train, 
                       method = "rf",
                       trControl = caret.train.control)

# Model details
machine.model

## Random Forest 
## 
## 17 samples
##  3 predictor
##  2 classes: '1', '2' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold, repeated 10 times) 
## Summary of sample sizes: 16, 16, 15, 15, 14, 15, ... 
## Resampling results across tuning parameters:
## 
##   mtry  Accuracy   Kappa    
##   2     0.8916667  0.7078125
##   3     0.8866667  0.6969231
##   5     0.8483333  0.6223881
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was mtry = 2.

The model, based on training data, is reporting a 89% accuracy. But the real test is on the test dataset, data.test:

# Predicting winners on test dataset
mm.predict <- predict(machine.model, newdata = data.test)

results <- cbind(data.test, mm.predict)

# Display results
kable(results, "html") %>%
        kable_styling(bootstrap_options = c("striped", "hover"))

The machine learning algorithm predicted that contestant 1 would win all 3 battles. It was right 2/3 times.

data.battle <- data.model[(nrow(data.model) - 1): nrow(data.model),]

# Machine prediction
mm.predict <- predict(machine.model, data.battle)

results <- cbind(data.battle, mm.predict)

# Display results
kable(results, "html") %>%
        kable_styling(bootstrap_options = c("striped", "hover"))