10-prediction.Rmd

# Prediction {#prediction-chapter}

```{r, child="_common.Rmd"}
```

<p style="font-weight:bold; color:red;">INCOMPLETE DRAFT</p>

> ...

```{block, type="rmdkey"}
The essential questions for this chapter are:

- ...
- ...
- ...
```

<!-- COURSE STRUCTURE

TUTORIALS:

- ...

SWIRL:

- ...

WORKED/ RECIPE:

- ...

PROJECT:

- ...

GOALS:

...

-->


<!-- IDEAS:

- English/ Spanish language detection (chapter)
  - ACTIVES
  - SUBTLEX_us
  - /Users/francojc/Documents/Academic/Teaching/Courses/2018-2019 Spring/LIN 380/Course/Video lectures/Language

- SPAM (recipe)
  - SMS messages
  - tadr::sms

- Nativeness (chapter)
  - CEDEL2 (Spanish)
  - Wricle and Locness (English?)
    - http://wricle.learnercorpora.com/
    - /Users/francojc/Documents/Academic/Research/Data/Language/Corpora/LOCNESS
      - Overall prediction native/ learner
      - ?Genitive construction markers native/ learner
      - ?Consider exploring the Mudpar approach

- Authorship (lab)
  - /Users/francojc/Documents/Academic/Research/Data/Language/Corpora/Federalist papers
  - /Users/francojc/Documents/Academic/Teaching/Courses/2018-2019 Spring/LIN 380/RStudio/Data/Federalist/Documents
  
- RateMyProfessor (from recipe 8)  
  - https://data.mendeley.com/datasets/fvtfjyvw7d/2

-->


```{r prediction-data-packages, include=FALSE}
# Packages
pacman::p_load(tidyverse, tidytext, caret, quanteda, quanteda.textstats, quanteda.textmodels)
```

In this chapter we will



Orientation to the question(s) and dataset(s) to be explored ...

```{r p-read-datasets}
wricle_df <- read_csv(file = "resources/10-prediction/data/derived/wricle_formal_curated.csv")

locness_df <- read_csv(file = "resources/10-prediction/data/derived/locness_curated.csv")
```

```{r}
learners_df <- 
  wricle_df %>% 
  filter(native_language == "Spanish") %>% 
  mutate(student = "Spanish") %>% 
  mutate(type = "Learner") %>% 
  select(essay_id = id, type, student, essay)

glimpse(learners_df)
```



```{r}
natives_df <- 
  locness_df %>% 
  unite(col = doc_id, c("file_id", "essay_id"), sep = "-") %>% 
  mutate(doc_id = str_replace(doc_id, "\\.txt", "")) %>% 
  mutate(type = "Native") %>% 
  select(essay_id = doc_id, type, student, essay)

glimpse(natives_df)
```

```{r}
nativeness_df <- rbind(learners_df, natives_df) # combine
```

```{r}
nativeness_df %>% 
  janitor::tabyl(type)
```

```{r}
nativeness_df %>% 
  unnest_tokens(output = "word", input = "essay") %>% 
  count(word, type) %>% 
  group_by(type) %>% 
  summarize(total_words = sum(n)) 
```
```{r}
nativeness_corpus <- 
  nativeness_df %>% 
  corpus(text_field = "essay")

nativeness_corpus_summary <- 
  nativeness_corpus %>% 
  summary(n = ndoc(nativeness_corpus))

nativeness_corpus_summary %>% 
  slice_head(n = 10)

```

```{r}
nativeness_corpus$doc_id <- 1:ndoc(nativeness_corpus)

nativeness_corpus %>% 
  docvars() %>% 
  slice_head(n = 5)
```

```{r}
nativeness_tokens <- 
  nativeness_corpus %>% 
  tokens(what = "word",
         remove_punct = TRUE, 
         remove_numbers = TRUE, 
         remove_symbols = TRUE)

nativeness_tokens %>% 
  tokens_group(groups = type) %>% 
  head()

```

```{r}
nativeness_dfm <- 
  nativeness_tokens %>% 
  dfm()

nativeness_dfm %>% 
  head(n = 5)

nativeness_dfm %>% 
  dfm_group(groups = type) %>% 
  head(n = 5)

```

```{r}
nativeness_dfm %>% 
  topfeatures(n = 25)
```

```{r}
nativeness_dfm %>% 
  textstat_frequency(n = 5, groups = type)
```

```{r}
nativeness_dfm %>% 
  dfm_tfidf() %>% 
  textstat_frequency(n = 5, groups = type, force = TRUE)
```
```{r}
set.seed(4321) # make reproducible

num_docs <- 
  nativeness_dfm %>% 
  ndoc()

train_size <- 
  (num_docs * .75) %>% # get size of sample
  round() # round to nearest whole number

train_ids <- sample(x = 1:num_docs, # population
                   size = train_size, # size of sample
                   replace = FALSE) # without replacement
```


```{r}
nativeness_dfm_train <- 
  nativeness_dfm %>% 
  dfm_subset(doc_id %in% train_ids)

nativeness_dfm_test <- 
  nativeness_dfm %>% 
  dfm_subset(!doc_id %in% train_ids)
```


```{r}
nativeness_dfm %>% 
  docvars() %>% 
  janitor::tabyl(type)

nativeness_dfm_train %>% 
  docvars() %>% 
  janitor::tabyl(type)

nativeness_dfm_test %>% 
  docvars() %>% 
  janitor::tabyl(type)
```


```{r}
nb1 <- textmodel_nb(x = nativeness_dfm_train, y = nativeness_dfm_train$type)

summary(nb1)
coef(nb1) %>% head()

predict(nb1, type = "prob") %>% # get the predicted document scores
  tail # preview predicted probability scores
```

```{r}
nb1_predictions <- 
  predict(nb1, type = "prob") %>% # get the predicted document scores
  as.data.frame() %>% # convert to data frame
  mutate(document = rownames(.)) %>% # add the document names to the data frame
  as_tibble() %>% # convert to tibble
  pivot_longer(cols = c("Learner", "Native"), # convert from wide to long format
               names_to = "prediction", # new column for ham/spam predictions
               values_to = "probability") %>% # probablity scores for each
  group_by(document) %>% # group parameter by document
  slice_max(probability, n = 1) %>% # keep the document row with highest probablity
  slice_head(n = 1) %>% # for predictions that were 50/50 
  ungroup() %>% # remove grouping parameter
  mutate(doc_id = str_remove(document, "text") %>% as.numeric) %>% # clean up document column so it matches doc_id in
  arrange(doc_id) # order by doc_id

nb1_predictions %>% 
  slice_head(n = 10) # preview
```


```{r}
nb1_predictions_actual <- 
  cbind(actual = nb1$y, nb1_predictions) %>% # column-bind actual classes
  select(doc_id, document, actual, prediction, probability) # organize variables

nb1_predictions_actual %>% 
  slice_head(n = 5) # preview
```

```{r}
tab_class <- 
  table(nb1_predictions_actual$actual, # actual class labels
        nb1_predictions_actual$prediction) # predicted class labels

caret::confusionMatrix(tab_class, mode = "prec_recall") # model performance statistics
```


```{r}
predicted_class <- 
  predict(nb1, newdata = nativeness_dfm_test)

actual_class <- nativeness_dfm_test$type

tab_class <- table(actual_class, predicted_class)  # cross-tabulate actual and predicted class labels

caret::confusionMatrix(tab_class, mode = "prec_recall")  # model performance statistics


```



## Preparation

Data set transformation

Splitting into training and test sets

## Model training

Aim to use the create an abstraction of the patterns in the dataset


- Feature engineering

- Model selection

- Model evaluation

The In Figure () Let's consider the results from a hypothetical model of text classification on the SMS dataset I introduced at in this subsection. 


```{r pda-contigency-table, echo=FALSE, eval=FALSE}
# Simulated table
# sms observations (testing 25%): 1210

knitr::include_graphics("images/10-prediction/pda-sms-contingency-table.png")

```

  - accuracy (measure of overall correct predictions)
  - precision (measure of the quality of the predictions)
    - Percentage of predicted 'ham' messages that were correct
  - recall (measure of the quantity of the predictions)
    - Percentage of actual 'ham' messages that were correct
  - F1-score (summarizes the balance between precision and recall)

Avoiding overfitting


## Model testing

Aim to test the abstracted model to new observations.  

Model testing

## Evaluation

Evaluation of results

Relationship between predicted and actual classes in a confusion matrix as seen in Figure \@ref(fig:pda-confusion-matrix-image).

```{r pda-confusion-matrix-image, echo=FALSE, fig.cap='Confusion matrix'}
knitr::include_graphics("images/10-prediction/pda-confusion-matrix.png")
```



## ...


In Table () we see the top five terms for each class after breaking the messages into terms and then counting up the frequencies.

```{r pred-sms-terms-freq, echo=FALSE, eval=FALSE}
sms_f <- 
  sms %>% 
  unnest_tokens(terms, message, token = "regex", pattern = " ") %>% 
  count(sms_type, terms) 

sms_f %>% 
  arrange(sms_type, desc(n)) %>% 
  group_by(sms_type) %>%
  slice_head(n = 5) %>% 
  select(sms_type, terms, frequency = n) %>% 
  kable(booktabs = TRUE,
        caption = "Top five most frequent terms for 'ham' and 'spam'.")
```

<!-- There are a couple things we may want to take into consideration given our first-pass results. First, there appears to some overlap in the top terms in each class (i.e. 'you', 'to', and 'a'). Ideally we want to create maximum separation between the terms that are most indicative of each class. Think of this from the point of view of a human learner. If you were to study a frequency list like this to try to learn how do distinguish between 'spam' and 'ham' and then you were given the term 'you' to make your decision on a test your decision would almost be a complete guess --the uncertainty is high because term figures prominently in both classes. Second you will note that the frequency for the first five terms of 'ham' is much higher than terms in 'spam'. This imbalance suggests that there are simply more 'ham' messages in our data or that 'ham' messages are longer (have more terms), or both. Any of these sources can potentially bias our machine learner to choose 'ham' over 'spam' for reasons which are not founded on the distinction between terms. -->

```{r pred-sms-dtm, echo=FALSE, eval=FALSE}
library(tidymodels)

sms_split <- initial_split(sms, strata = "sms_type")
train_data <- training(sms_split)
test_data <- testing(sms_split)

library(textrecipes)
```



## Summary

...