Toxic Comment Filter

1 Introduction

Build a model that can filter user comments based on the degree of language maliciousness:

• Preprocess the text by eliminating the set of tokens that do not make significant contribution at the semantic level.
• Transform the text corpus into sequences.
• Build a Deep Learning model including recurrent layers for a multilabel classification task.
• At prediction time, the model should return a vector containing a 1 or a 0 at each label in the dataset (toxic, severe_toxic, obscene, threat, insult, identity_hate). In this way, a non-harmful comment will be classified by a vector of only 0s [0,0,0,0,0]. In contrast, a dangerous comment will exhibit at least a 1 among the 6 labels.

2 Setup

Leveraging Quarto and RStudio, I will setup an R and Python enviroment.

2.1 Import R libraries

Import R libraries. These will be used for both the rendering of the document and data analysis. The reason is I prefer ggplot2 over matplotlib. I will also use colorblind safe palettes.

library(tidyverse, verbose = FALSE)
library(tidymodels, verbose = FALSE)
library(reticulate)
library(ggplot2)
library(plotly)
library(RColorBrewer)
library(bslib)
library(Metrics)
library(gt)

reticulate::use_virtualenv("r-tf")

2.2 Import Python packages

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
import keras
import keras_nlp

from keras.backend import clear_session
from keras.models import Model, load_model
from keras.layers import TextVectorization, Input, Dense, Embedding, Dropout, GlobalAveragePooling1D, LSTM, Bidirectional, GlobalMaxPool1D, Flatten, Attention
from keras.metrics import Precision, Recall, AUC, SensitivityAtSpecificity, SpecificityAtSensitivity, F1Score


from sklearn.model_selection import train_test_split, KFold
from sklearn.metrics import multilabel_confusion_matrix, classification_report, ConfusionMatrixDisplay, precision_recall_curve, f1_score, recall_score, roc_auc_score

Create a Config class to store all the useful parameters for the model and for the project.

2.3 Class Config

I created a class with all the basic configuration of the model, to improve the readability.

class Config():
    def __init__(self):
        self.url = "https://s3.eu-west-3.amazonaws.com/profession.ai/datasets/Filter_Toxic_Comments_dataset.csv"
        self.max_tokens = 20000
        self.output_sequence_length = 911 # check the analysis done to establish this value
        self.embedding_dim = 128
        self.batch_size = 32
        self.epochs = 100
        self.temp_split = 0.3
        self.test_split = 0.5
        self.random_state = 42
        self.total_samples = 159571 # total train samples
        self.train_samples = 111699
        self.val_samples = 23936
        self.features = 'comment_text'
        self.labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']
        self.new_labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate', "clean"]
        self.label_mapping = {label: i for i, label in enumerate(self.labels)}
        self.new_label_mapping = {label: i for i, label in enumerate(self.labels)}
        self.path = "/Users/simonebrazzi/R/blog/posts/toxic_comment_filter/history/f1score/"
        self.model =  self.path + "model_f1.keras"
        self.checkpoint = self.path + "checkpoint.lstm_model_f1.keras"
        self.history = self.path + "lstm_model_f1.xlsx"
        
        self.metrics = [
            Precision(name='precision'),
            Recall(name='recall'),
            AUC(name='auc', multi_label=True, num_labels=len(self.labels)),
            F1Score(name="f1", average="macro")
            
        ]
    def get_early_stopping(self):
        early_stopping = keras.callbacks.EarlyStopping(
            monitor="val_f1", # "val_recall",
            min_delta=0.2,
            patience=10,
            verbose=0,
            mode="max",
            restore_best_weights=True,
            start_from_epoch=3
        )
        return early_stopping

    def get_model_checkpoint(self, filepath):
        model_checkpoint = keras.callbacks.ModelCheckpoint(
            filepath=filepath,
            monitor="val_f1", # "val_recall",
            verbose=0,
            save_best_only=True,
            save_weights_only=False,
            mode="max",
            save_freq="epoch"
        )
        return model_checkpoint

    def find_optimal_threshold_cv(self, ytrue, yproba, metric, thresholds=np.arange(.05, .35, .05), n_splits=7):

      # instantiate KFold
      kf = KFold(n_splits=n_splits, shuffle=True, random_state=42)
      threshold_scores = []

      for threshold in thresholds:

        cv_scores = []
        for train_index, val_index in kf.split(ytrue):

          ytrue_val = ytrue[val_index]
          yproba_val = yproba[val_index]

          ypred_val = (yproba_val >= threshold).astype(int)
          score = metric(ytrue_val, ypred_val, average="macro")
          cv_scores.append(score)

        mean_score = np.mean(cv_scores)
        threshold_scores.append((threshold, mean_score))

        # Find the threshold with the highest mean score
        best_threshold, best_score = max(threshold_scores, key=lambda x: x[1])
      return best_threshold, best_score

config = Config()

3 Data

The dataset is accessible using tf.keras.utils.get_file to get the file from the url. N.B. For reproducibility purpose, I also downloaded the dataset. There was time in which the link was not available.

# df = pd.read_csv(config.path)
file = tf.keras.utils.get_file("Filter_Toxic_Comments_dataset.csv", config.url)
df = pd.read_csv(file)

library(reticulate)

py$df %>%
  tibble() %>% 
  head(5) %>% 
  gt() %>% 
  tab_header(
    title = "First five observations"
  ) %>% 
   cols_align(
    align = "center",
    columns = c("toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate", "sum_injurious")
  ) %>% 
  cols_align(
    align = "left",
    columns = comment_text
  ) %>% 
  cols_label(
    comment_text = "Comments",
    toxic = "Toxic",
    severe_toxic = "Severe Toxic",
    obscene = "Obscene",
    threat = "Threat",
    insult = "Insult",
    identity_hate = "Identity Hate",
    sum_injurious = "Sum Injurious"
    )

Table 1: First 5 elemtns

First five observations
Comments	Toxic	Severe Toxic	Obscene	Threat	Insult	Identity Hate	Sum Injurious
Explanation Why the edits made under my username Hardcore Metallica Fan were reverted? They weren't vandalisms, just closure on some GAs after I voted at New York Dolls FAC. And please don't remove the template from the talk page since I'm retired now.89.205.38.27	0	0	0	0	0	0	0
D'aww! He matches this background colour I'm seemingly stuck with. Thanks. (talk) 21:51, January 11, 2016 (UTC)	0	0	0	0	0	0	0
Hey man, I'm really not trying to edit war. It's just that this guy is constantly removing relevant information and talking to me through edits instead of my talk page. He seems to care more about the formatting than the actual info.	0	0	0	0	0	0	0
" More I can't make any real suggestions on improvement - I wondered if the section statistics should be later on, or a subsection of ""types of accidents"" -I think the references may need tidying so that they are all in the exact same format ie date format etc. I can do that later on, if no-one else does first - if you have any preferences for formatting style on references or want to do it yourself please let me know. There appears to be a backlog on articles for review so I guess there may be a delay until a reviewer turns up. It's listed in the relevant form eg Wikipedia:Good_article_nominations#Transport "	0	0	0	0	0	0	0
You, sir, are my hero. Any chance you remember what page that's on?	0	0	0	0	0	0	0

Lets create a clean variable for EDA purpose: I want to visually see how many observation are clean vs the others labels.

df.loc[df.sum_injurious == 0, "clean"] = 1
df.loc[df.sum_injurious != 0, "clean"] = 0

3.1 EDA

First a check on the dataset to find possible missing values and imbalances.

3.1.1 Frequency

library(reticulate)
df_r <- py$df
new_labels_r <- py$config$new_labels

df_r_grouped <- df_r %>% 
  select(all_of(new_labels_r)) %>%
  pivot_longer(
    cols = all_of(new_labels_r),
    names_to = "label",
    values_to = "value"
  ) %>% 
  group_by(label) %>%
  summarise(count = sum(value)) %>% 
  mutate(freq = round(count / sum(count), 4))

df_r_grouped %>% 
  gt() %>% 
  tab_header(
    title = "Labels frequency",
    subtitle = "Absolute and relative frequency"
  ) %>% 
  fmt_number(
    columns = "count",
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = TRUE,
    use_seps = TRUE
  ) %>% 
  fmt_percent(
    columns = "freq",
    decimals = 2,
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = FALSE
  ) %>% 
  cols_align(
    align = "center",
    columns = c("count", "freq")
  ) %>% 
  cols_align(
    align = "left",
    columns = label
  ) %>% 
  cols_label(
    label = "Label",
    count = "Absolute Frequency",
    freq = "Relative frequency"
  )

Table 2: Absolute and relative labels frequency

Labels frequency
Absolute and relative frequency
Label	Absolute Frequency	Relative frequency
clean	143,346	80.33%
identity_hate	1,405	0.79%
insult	7,877	4.41%
obscene	8,449	4.73%
severe_toxic	1,595	0.89%
threat	478	0.27%
toxic	15,294	8.57%

3.1.2 Barchart

library(reticulate)
barchart <- df_r_grouped %>%
  ggplot(aes(x = reorder(label, count), y = count, fill = label)) +
  geom_col() +
  labs(
    x = "Labels",
    y = "Count"
  ) +
  # sort bars in descending order
  scale_x_discrete(limits = df_r_grouped$label[order(df_r_grouped$count, decreasing = TRUE)]) +
  scale_fill_brewer(type = "seq", palette = "RdYlBu") +
  theme_minimal()
ggplotly(barchart)

Figure 1: Imbalance in the dataset with clean variable

It is visible how much the dataset in imbalanced. This means it could be useful to check for the class weight and use this argument during the training.

It is clear that most of our text are clean. We are talking about 0.8033 of the observations which are clean. Only 0.1967 are toxic comments.

3.2 Sequence lenght definition

To convert the text in a useful input for a NN, it is necessary to use a TextVectorization layer. See the Section 4 section.

One of the method is output_sequence_length: to better define it, it is useful to analyze our text length. To simulate what the model we do, we are going to remove the punctuation and the new lines from the comments.

3.2.1 Summary

library(reticulate)
df_r %>% 
  mutate(
    comment_text_clean = comment_text %>%
      tolower() %>% 
      str_remove_all("[[:punct:]]") %>% 
      str_replace_all("\n", " "),
    text_length = comment_text_clean %>% str_count()
    ) %>% 
  pull(text_length) %>% 
  summary() %>% 
  as.list() %>% 
  as_tibble() %>% 
  gt() %>% 
  tab_header(
    title = "Summary Statistics",
    subtitle = "of text length"
  ) %>% 
  fmt_number(
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = TRUE,
    use_seps = TRUE
  ) %>% 
  cols_align(
    align = "center",
  ) %>% 
  cols_label(
    Min. = "Min",
    `1st Qu.` = "Q1",
    Median = "Median",
    `3rd Qu.` = "Q3",
    Max. = "Max"
  )

Table 3: Summary of text length

Summary Statistics
of text length
Min	Q1	Median	Mean	Q3	Max
4	91	196	378.4	419	5,000

3.2.2 Boxplot

library(reticulate)
boxplot <- df_r %>% 
  mutate(
    comment_text_clean = comment_text %>%
      tolower() %>% 
      str_remove_all("[[:punct:]]") %>% 
      str_replace_all("\n", " "),
    text_length = comment_text_clean %>% str_count()
    ) %>% 
  # pull(text_length) %>% 
  ggplot(aes(y = text_length)) +
  geom_boxplot() +
  theme_minimal()
ggplotly(boxplot)

Figure 2: Text length boxplot

3.2.3 Histogram

library(reticulate)
df_ <- df_r %>% 
  mutate(
    comment_text_clean = comment_text %>%
      tolower() %>% 
      str_remove_all("[[:punct:]]") %>% 
      str_replace_all("\n", " "),
    text_length = comment_text_clean %>% str_count()
  )

Q1 <- quantile(df_$text_length, 0.25)
Q3 <- quantile(df_$text_length, 0.75)
IQR <- Q3 - Q1
upper_fence <- as.integer(Q3 + 1.5 * IQR)

histogram <- df_ %>% 
  ggplot(aes(x = text_length)) +
  geom_histogram(bins = 50) +
  geom_vline(aes(xintercept = upper_fence), color = "red", linetype = "dashed", linewidth = 1) +
  theme_minimal() +
  xlab("Text Length") +
  ylab("Frequency") +
  xlim(0, max(df_$text_length, upper_fence))
ggplotly(histogram)

Figure 3: Text length histogram with boxplot upper fence

Considering all the above analysis, I think a good starting value for the output_sequence_length is 911, the upper fence of the boxplot. In the last plot, it is the dashed red vertical line.. Doing so, we are removing the outliers, which are a small part of our dataset.

3.3 Dataset

Now we can split the dataset in 3: train, test and validation sets. Considering there is not a function in sklearn which lets split in these 3 sets, we can do the following: - split between a train and temporary set with a 0.3 split. - split the temporary set in 2 equal sized test and val sets.

x = df[config.features].values
y = df[config.labels].values

xtrain, xtemp, ytrain, ytemp = train_test_split(
  x,
  y,
  test_size=config.temp_split, # .3
  random_state=config.random_state
  )
xtest, xval, ytest, yval = train_test_split(
  xtemp,
  ytemp,
  test_size=config.test_split, # .5
  random_state=config.random_state
  )

xtrain shape: py$xtrain.shape ytrain shape: py$ytrain.shape xtest shape: py$xtest.shape ytest shape: py$ytest.shape xval shape: py$xval.shape yval shape: py$yval.shape

The datasets are created using the tf.data.Dataset function. It creates a data input pipeline. The tf.data API makes it possible to handle large amounts of data, read from different data formats, and perform complex transformations. The tf.data.Dataset is an abstraction that represents a sequence of elements, in which each element consists of one or more components. Here each dataset is creates using from_tensor_slices. It create a tf.data.Dataset from a tuple (features, labels). .batch let us work in batches to improve performance, while .prefetch overlaps the preprocessing and model execution of a training step. While the model is executing training step s, the input pipeline is reading the data for step s+1. Check the documentation for further informations.

train_ds = (
    tf.data.Dataset
    .from_tensor_slices((xtrain, ytrain))
    .shuffle(xtrain.shape[0])
    .batch(config.batch_size)
    .prefetch(tf.data.experimental.AUTOTUNE)
)

test_ds = (
    tf.data.Dataset
    .from_tensor_slices((xtest, ytest))
    .batch(config.batch_size)
    .prefetch(tf.data.experimental.AUTOTUNE)
)

val_ds = (
    tf.data.Dataset
    .from_tensor_slices((xval, yval))
    .batch(config.batch_size)
    .prefetch(tf.data.experimental.AUTOTUNE)
)

print(
  f"train_ds cardinality: {train_ds.cardinality()}\n",
  f"val_ds cardinality: {val_ds.cardinality()}\n",
  f"test_ds cardinality: {test_ds.cardinality()}\n"
  )

train_ds cardinality: 3491
 val_ds cardinality: 748
 test_ds cardinality: 748

Check the first element of the dataset to be sure that the preprocessing is done correctly.

train_ds.as_numpy_iterator().next()

(array([b'"\nThanks, a separate sentence is an excellent idea, but we will need reliable sources for it. Rather, than tag bomb the sentence, I\'ll do it here:\n""typically in education""\n""everybody knows"" (students are included in ""everybody"", but they don\'t know ""the other kind of numbers"")\n""natural numbers  counting numbers"" (natural numbers with or without zero?)\nWhy was the term ""counting numbers"" introduced ""circa 1965""? (per Merriam-webster.com)\nAt what educational level is the term ""counting numbers"" replaced with another term?\nSourcing ideas include books on pre-algebra and mathematics curriculum standards.\n50.53.39.110  "',
       b',Wikipedia:Articles for deletion/List of countries by ratio of carbon dioxide emissions to GDP per capita',
       b"It doesn't like them, but I thought the problem had been fixed. I'll run some more unit tests.",
       b"Amendment of template \n\nIs there any reason the Obama template can't be amended with a controversies line, on which we would mention Wright, Ayres, and birthplace?",
       b'"\nI concur. Outrageous breach of natural justice by an admin....(kiwiexile at DMOZ)  \xe2\x80\xa2 \n\n"',
       b'"\nAnd the entirety of your edits are not related to feminist topics? What a weird comment considering you represent the opposing viewpoint of Memills but still somehow represent this like you\'d be somehow more neutral. In your earlier comment on this page, you said that Memills ""subscribes to a somewhat controversial branch of evolutionary psychology"". Well, may I ask, do you subscribe to the somewhat controversial branch women\'s studies of sociology?\nRegardless, I suggest everyone reads the following section on KillerChihuahua\'s talk page: . KillerChihuahua has been giving ""admin support"" to Slp1\'s side as pointed out themselves. I suggest you Memills fill a complaint on this. \' "',
       b'"\n\n Your message to me \n\nThe page you linked to said ""by convention your user page will usually not be edited by others"". 122.105.217.56  "',
       b'And I see what you meant now, too. -) I completely missed that he changed font sizes. About the discussion, I believe I prefer the 100% standard version as well. Hopefully it will make for less edit warring and a cleaner Wikipedia.  hideliho!',
       b'Gmaxwell\nHi, he has said he does want the image here. Cheers',
       b'"\n\nThis message is regarding the page User:Jpgordon. Please stop.  If you continue to vandalize pages, you will be blocked from editing Wikipedia.  \xc2\xab\xc2\xbb?\xc2\xbf?meta "',
       b'"::::::::::::::I would suggest that it\'s not necessarily an ""anti-vax"" site as suggested, and it does cite sources within the body if was read.  Nonetheless, it raises interesting points (such as efficacy being related to age and to a particular strain)and does present the other side of the argument.  With respect to purported anti-vax statements on websites, this is against the standard of practice in most Canadian provinces and the CCO has fined members up to 25K for those shenanigans.  Nonetheless, we should not pick one example and make attributions, for I can go to MD websites and show them endorsing questionable methods such as homeopathy.  We need to be careful with giving fringe POVs too much weight here which seems to be an increasing concern.   \n"',
       b'"\n\nI can say the furlong is a known length in the US and it even sees use on the interstate system as 1/8 mile increments are occasionally signed (although I have never seen 1/8 mile, think it was 3/8 ad 5/8 that I saw).  I\'ve seen several old British signs that are still in place along roadsigns of public roads in the UK that a buddy of mine there showed me.  Perhaps if he or the author consents, I could put them up here?  Think one was furlongs and there may or may not have been one that was in miles and eighths.\n\nAlso, does anyone have some information about the Myanmar furlong sign.  Some internet armchair expert was claiming they were all kilometers ""now"" which I had assumed correct, until seeing an even newer post, less than 6 mos. old talking about biking there and how ""everything was still in miles [and furlongs] there.  If we had a more current picture of the signage there, even if it\'s the same damned toll plaza, that would go a long way.  Their government website, at least in English, is, shall we say, lacking?\n  "',
       b"OK, but when you threaten to lock a page because I don't agree with your recommendations, I think calling you a dictator is more of an observation rather than a personal attack.",
       b'The problem is no longer there.',
       b'Thank you for experimenting with Wikipedia. Your test worked, and it has been reverted or removed. Please use the sandbox for any other tests you want to do. Take a look at the welcome page if you would like to learn more about contributing to our encyclopedia.    | Talk',
       b'"\n\nWelcome\n\nHello and welcome to Wikipedia!  We appreciate encyclopedic contributions, but some of your recent contributions do not conform to our policies.  For more information on this, see\nWikipedia\'s policies on vandalism\nPolicies on banned or discouraged content\n\nIf you\'d like to experiment with the wiki\'s syntax, please do so on Wikipedia:Sandbox rather than in articles.\n\nIf you still have questions, there is a new contributor\'s help page, or you can write {{helpme}} below this message along with a question and someone will be along to answer it shortly.  You may also find the following pages useful for a general introduction to Wikipedia.\nThe five pillars of Wikipedia\nHow to edit a page\nHelp pages\nTutorial\nHow to write a great article\nManual of Style\nPolicy on neutral point of view\nGuideline on external links\nGuideline on conflict of interest\nI hope you enjoy editing Wikipedia!  Please sign your name on talk pages using four tildes (~~~~); this will automatically produce your name and the date. Feel free to write a note on the bottom of  if you want to get in touch with me. Again, welcome!  - [[User_talk:Bpeps|t@lk]] "',
       b'oh i like that one you ask me on the page to state my case i do so and you erase it - in debate was i was a little child i learned that you just forfeited 68.231.15.56',
       b'Can you unblock my IP? \n\nHey pgk, can you plz unblock my ip, cos i have been blocked cos sum1 else was blocked using my AOL IP. Thanks,',
       b"Usually, I'd agree, but for two reasons I think not in this case.... A- it would look choppy, and B- it is still the MAIN article, since it was the most pathetic example.  This is, however, just my opinion.  BTW dont sweat it.  I always forget to sign in (KA-BOOOOM!!!!)",
       b'Please stop adding nonsense to Wikipedia. It is considered vandalism. If you were just trying to experiment, then use the sandbox instead. Thank you.',
       b', 20 September 2005 (UTC)\n\n ... or (C) their loved one is brought back to life. (B) is by far the most likely though.   |  Talk 00:38',
       b'Punk\n\nthe only reason anyone gives you  trouble is because you cause trouble yourself. You deserve to be attacked',
       b'. \n\nP.s. Victor actually told me himself at the practice. What do I have to do to prove it? Record him?!? LOL yea son',
       b'"\nI do not see any anti-semitism here. If the guy gives lectures in New York synagogues and knocks down his country every single minute, using every opportunity in such influential newspapers as New York Post, then one can state that his interests are not only personal. You can not find any intellectual person among neoconservatives in New York Post. Michelle Malkin, Andrea Peyser, Deborah Orin (may her rest in peace, although I know she is rotting in hell)and other essayists are the ones who make a great company for Amir Taheri. This guy has no mind at all. I agree that Iran\'s regime is wrong, but the regime of Shah that Amir Taheri is longing for, because he owned Keyhan during that regime was not any better and Islamic Revolution was justified. Look at how they intriduce Amir Taheri in both this article and in New York Post. ""Based in Europe memeber of Benador Associates"". Is Europe that small? Is that a small town? Why don\'t they say based in the world journalist. Or may be he is having a breakfast in London with Christian Amanpour and James Rubin, eating lunch in Jerusalem, and having some lectures and dinner with anti Irani Jews in Forest Hills, New York in between emailing his full of hatred articles to New York Post. Isn\'t this stupid? How can you saysomebody is based in Europe or Asia or Africa?"',
       b'Wikileaks article  is written in British English, and some terms used in it are different or absent from American English and other varieties of English. According to the relevant style guide, this should not be changed without broad consensus.',
       b'"\n\ncriticism section\n\nPeople have repeatedly tried to add a criticism section.. and it keeps getting deleted.. If anyone warrents a criticism secion it is alex jones with his nutty theories of black helicopters,secret government detention facilities at amtrak stations retrofitted with gas chambers for american citizens, and so on and so on.. nevermind the self proclaimed crusader for truth admits on his website that you can buy advertisements ""disguised as news headlines"" in the the police state section.  real thruthfull.. I mean seriously.. an admin needs to add this, ban people who keep censoring the article the protect the page if necessary... this is rediculous. -  "',
       b"I am trying to update my OWN page and you keep undoing it and then accusing me of adding uncourced content, IAM the SOURCE what can I dio? verty fryustrating, please can you leave my page alone, i'm trying to keep my list of publications up to date, and don't block me \nChristopher Hart",
       b'Please stop. If you continue to blank out or delete portions of page content, templates or other materials from Wikipedia, you will be blocked from editing.  -38.116.200.85',
       b'Hello \n\nPlease u deleted the page on Patrick Obahiagbon. I really dont know why although I cited my references and it has be categorised into Nigerian Politicians by another admin. Please why did u delete it? *sad*',
       b"The Comedy of Errors\nSorry, but I've reverted your additions to The Comedy of Errors, as they don't really belong there. Sure, they're about Shakespeare, but you haven't shown how they're in any way relevant to he article about that play, and they don't quite fit in. Thanks.",
       b'You are IGNORING my other sources, and only referring to my forum source. Why not put a tag there so I can fix it myself instead of WRECKING hours of work.',
       b'alreajk not logging in.'], dtype=object), array([[0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [1, 0, 0, 0, 1, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0]]))

And we check also the shape. We expect a feature of shape (batch, ) and a target of shape (batch, number of labels).

print(
  f"text train shape: {train_ds.as_numpy_iterator().next()[0].shape}\n",
  f" text train type: {train_ds.as_numpy_iterator().next()[0].dtype}\n",
  f"label train shape: {train_ds.as_numpy_iterator().next()[1].shape}\n",
  f"label train type: {train_ds.as_numpy_iterator().next()[1].dtype}\n"
  )

text train shape: (32,)
  text train type: object
 label train shape: (32, 6)
 label train type: int64

4 Preprocessing

Of course preprocessing! Text is not the type of input a NN can handle. The TextVectorization layer is meant to handle natural language inputs. The processing of each example contains the following steps: 1. Standardize each example (usually lowercasing + punctuation stripping) 2. Split each example into substrings (usually words) 3. Recombine substrings into tokens (usually ngrams) 4. Index tokens (associate a unique int value with each token) 5. Transform each example using this index, either into a vector of ints or a dense float vector.

For more reference, see the documentation at the following link.

text_vectorization = TextVectorization(
  max_tokens=config.max_tokens,
  standardize="lower_and_strip_punctuation",
  split="whitespace",
  output_mode="int",
  output_sequence_length=config.output_sequence_length,
  pad_to_max_tokens=True
  )

# prepare a dataset that only yields raw text inputs (no labels)
text_train_ds = train_ds.map(lambda x, y: x)
# adapt the text vectorization layer to the text data to index the dataset vocabulary
text_vectorization.adapt(text_train_ds)

This layer is set to: - max_tokens: 20000. It is common for text classification. It is the maximum size of the vocabulary for this layer. - output_sequence_length: 911. See Figure 3 for the reason why. Only valid in "int" mode. - output_mode: outputs integer indices, one integer index per split string token. When output_mode == “int”, 0 is reserved for masked locations; this reduces the vocab size to max_tokens - 2 instead of max_tokens - 1. - standardize: "lower_and_strip_punctuation". - split: on whitespace.

To preserve the original comments as text and also have a tf.data.Dataset in which the text is preprocessed by the TextVectorization function, it is possible to map it to the features of each dataset.

processed_train_ds = train_ds.map(
    lambda x, y: (text_vectorization(x), y),
    num_parallel_calls=tf.data.experimental.AUTOTUNE
)
processed_val_ds = val_ds.map(
    lambda x, y: (text_vectorization(x), y),
    num_parallel_calls=tf.data.experimental.AUTOTUNE
)
processed_test_ds = test_ds.map(
    lambda x, y: (text_vectorization(x), y),
    num_parallel_calls=tf.data.experimental.AUTOTUNE
)

5 Model

5.1 Definition

Define the model using the Functional API.

def get_deeper_lstm_model():
    clear_session()
    inputs = Input(shape=(None,), dtype=tf.int64, name="inputs")
    embedding = Embedding(
        input_dim=config.max_tokens,
        output_dim=config.embedding_dim,
        mask_zero=True,
        name="embedding"
    )(inputs)
    x = Bidirectional(LSTM(256, return_sequences=True, name="bilstm_1"))(embedding)
    x = Bidirectional(LSTM(128, return_sequences=True, name="bilstm_2"))(x)
    # Global average pooling
    x = GlobalAveragePooling1D()(x)
    # Add regularization
    x = Dropout(0.3)(x)
    x = Dense(64, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.01))(x)
    x = LayerNormalization()(x)
    outputs = Dense(len(config.labels), activation='sigmoid', name="outputs")(x)
    model = Model(inputs, outputs)
    model.compile(optimizer='adam', loss="binary_crossentropy", metrics=config.metrics, steps_per_execution=32)
    
    return model

lstm_model = get_deeper_lstm_model()
lstm_model.summary()

5.2 Callbacks

Finally, the model has been trained using 2 callbacks: - Early Stopping, to avoid to consume the kaggle GPU time. - Model Checkpoint, to retrieve the best model training information.

my_es = config.get_early_stopping()
my_mc = config.get_model_checkpoint(filepath="/checkpoint.keras")
callbacks = [my_es, my_mc]

5.3 Final preparation before fit

Considering the dataset is imbalanced, to increase the performance we need to calculate the class weight. This will be passed during the training of the model.

lab = pd.DataFrame(columns=config.labels, data=ytrain)
r = lab.sum() / len(ytrain)
class_weight = dict(zip(range(len(config.labels)), r))
df_class_weight = pd.DataFrame.from_dict(
  data=class_weight,
  orient='index',
  columns=['class_weight']
  )
df_class_weight.index = config.labels

library(reticulate)
py$df_class_weight %>% 
  gt() %>% 
  fmt_percent(
    decimals = 2,
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = TRUE
  )

Table 4: Class weight

class_weight
9.59%
0.99%
5.28%
0.31%
4.91%
0.87%

It is also useful to define the steps per epoch for train and validation dataset. This step is required to avoid to not consume entirely the dataset during the fit, which happened to me.

steps_per_epoch = config.train_samples // config.batch_size
validation_steps = config.val_samples // config.batch_size

5.4 Fit

The fit has been done on Kaggle to levarage the GPU. Some considerations about the model:

• .repeat() ensure the model sees all the dataset.
• epocs is set to 100.
• validation_data has the same repeat.
• callbacks are the one defined before.
• class_weight ensure the model is trained using the frequency of each class, because our dataset is imbalanced.
• steps_per_epoch and validation_steps depend on the use of repeat.

history = model.fit(
  processed_train_ds.repeat(),
  epochs=config.epochs,
  validation_data=processed_val_ds.repeat(),
  callbacks=callbacks,
  class_weight=class_weight,
  steps_per_epoch=steps_per_epoch,
  validation_steps=validation_steps
  )

Now we can import the model and the history trained on Kaggle.

model = load_model(filepath=config.model)
history = pd.read_excel(config.history)

5.5 Evaluate

validation = model.evaluate(
  processed_val_ds.repeat(),
  steps=validation_steps, # 748
  verbose=0
  )

val_metrics <- tibble(
  metric = c("loss", "precision", "recall", "auc", "f1_score"),
  value = py$validation
  )
val_metrics %>% 
  gt() %>% 
  fmt_number(
    columns = c("value"),
    decimals = 4,
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = TRUE
  ) %>% 
  cols_align(
    align = "left",
    columns = metric
  ) %>% 
  cols_align(
    align = "center",
    columns = value
  ) %>% 
  cols_label(
    metric = "Metric",
    value = "Value"
  )

Table 5: Model validation metric

Metric	Value
loss	0.0542
precision	0.7888
recall	0.671
auc	0.9572
f1_score	0.0293

5.6 Predict

For the prediction, the model does not need to repeat the dataset, because it has already been trained on all of the train data. Now it has just to consume the new data to make the prediction.

predictions = model.predict(processed_test_ds, verbose=0)

5.7 Confusion Matrix

The best way to assess the performance of a multi label classification is using a confusion matrix. Sklearn has a specific function to create a multi label classification matrix to handle the fact that there could be multiple labels for one prediction.

5.7.1 Grid Search Cross Validation for best threshold

Grid Search CV is a technique for fine-tuning hyperparameter of a ML model. It systematically search through a set of hyperparamenter values to find the combination which led to the best model performance. In this case, I am using a KFold Cross Validation is a resempling technique to split the data into k consecutive folds. Each fold is used once as a validation while the k - 1 remaining folds are the training set. See the documentation for more information.

The model is trained to optimize the recall. The decision was made because the cost of missing a True Positive is greater than a False Positive. In this case, missing a injurious observation is worst than classifying a clean one as bad.

5.7.2 Confidence threshold and Precision-Recall trade off

Whilst the KFold GDCV technique is usefull to test multiple hyperparameter, it is important to understand the problem we are facing. A multi label deep learning classifier outputs a vector of per-class probabilities. These need to be converted to a binary vector using a confidence threshold.

• The higher the threshold, the less classes the model predicts, increasing model confidence [higher Precision] and increasing missed classes [lower Recall].
• The lower the threshold, the more classes the model predicts, decreasing model confidence [lower Precision] and decreasing missed classes [higher Recall].

Threshold selection mean we have to decide which metric to prioritize, based on the problem we are facing and the relative cost of misduging. We can consider the toxic comment filtering a problem similiar to cancer diagnostic. It is better to predict cancer in people who do not have it [False Positive] and perform further analysis than do not predict cancer when the patient has the disease [False Negative].

I decide to train the model on the F1 score to have a balanced model in both precision and recall and leave to the threshold selection to increase the recall performance.

Moreover, the model has been trained on the macro avarage F1 score, which is a single performance indicator obtained by the mean of the Precision and Recall scores of individual classses.

\[ F1\ macro\ avg = \frac{\sum_{i=1}^{n} F1_i}{n} \]

It is useful with imbalanced classes, because it weights each classes equally. It is not influenced by the number of samples of each classes. This is sette both in the config.metrics and find_optimal_threshold_cv.

f1_score

ytrue = ytest.astype(int)
y_pred_proba = predictions
optimal_threshold_f1, best_score_f1 = config.find_optimal_threshold_cv(ytrue, y_pred_proba, f1_score)

print(f"Optimal threshold: {optimal_threshold_f1}")

Optimal threshold: 0.15000000000000002

print(f"Best score: {best_score_f1}")

Best score: 0.4788653077945807

# Use the optimal threshold to make predictions
final_predictions_f1 = (y_pred_proba >= optimal_threshold_f1).astype(int)

Optimal threshold f1 score: 0.15. Best score: 0.4788653.

recall_score

ytrue = ytest.astype(int)
y_pred_proba = predictions
optimal_threshold_recall, best_score_recall = config.find_optimal_threshold_cv(ytrue, y_pred_proba, recall_score)

# Use the optimal threshold to make predictions
final_predictions_recall = (y_pred_proba >= optimal_threshold_recall).astype(int)

Optimal threshold recall: 0.05. Best score: 0.8095814.

roc_auc_score

ytrue = ytest.astype(int)
y_pred_proba = predictions
optimal_threshold_roc, best_score_roc = config.find_optimal_threshold_cv(ytrue, y_pred_proba, roc_auc_score)

print(f"Optimal threshold: {optimal_threshold_roc}")

Optimal threshold: 0.05

print(f"Best score: {best_score_roc}")

Best score: 0.8809499649742268

# Use the optimal threshold to make predictions
final_predictions_roc = (y_pred_proba >= optimal_threshold_roc).astype(int)

Optimal threshold roc: 0.05. Best score: 0.88095.

5.7.3 Confusion Matrix Plot

# convert probability predictions to predictions
ypred = predictions >=  optimal_threshold_recall # .05
ypred = ypred.astype(int)

# create a plot with 3 by 2 subplots
fig, axes = plt.subplots(3, 2, figsize=(15, 15))
axes = axes.flatten()
mcm = multilabel_confusion_matrix(ytrue, ypred)
# plot the confusion matrices for each label
for i, (cm, label) in enumerate(zip(mcm, config.labels)):
    disp = ConfusionMatrixDisplay(confusion_matrix=cm)
    disp.plot(ax=axes[i], colorbar=False)
    axes[i].set_title(f"Confusion matrix for label: {label}")
plt.tight_layout()
plt.show()

Figure 4: Multi Label Confusion matrix

5.8 Classification Report

cr = classification_report(
  ytrue,
  ypred,
  target_names=config.labels,
  digits=4,
  output_dict=True
  )
df_cr = pd.DataFrame.from_dict(cr).reset_index()

library(reticulate)
df_cr <- py$df_cr %>% dplyr::rename(names = index)
cols <- df_cr %>% colnames()
df_cr %>% 
  pivot_longer(
    cols = -names,
    names_to = "metrics",
    values_to = "values"
  ) %>% 
  pivot_wider(
    names_from = names,
    values_from = values
  ) %>% 
  gt() %>%
  tab_header(
    title = "Confusion Matrix",
    subtitle = "Threshold optimization favoring recall"
  ) %>% 
  fmt_number(
    columns = c("precision", "recall", "f1-score", "support"),
    decimals = 2,
    drop_trailing_zeros = TRUE,
    drop_trailing_dec_mark = FALSE
  ) %>% 
  cols_align(
    align = "center",
    columns = c("precision", "recall", "f1-score", "support")
  ) %>% 
  cols_align(
    align = "left",
    columns = metrics
  ) %>% 
  cols_label(
    metrics = "Metrics",
    precision = "Precision",
    recall = "Recall",
    `f1-score` = "F1-Score",
    support = "Support"
  )

Table 6: Classification report

Confusion Matrix
Threshold optimization favoring recall
Metrics	Precision	Recall	F1-Score	Support
toxic	0.55	0.89	0.68	2,262.
severe_toxic	0.24	0.92	0.37	240.
obscene	0.55	0.94	0.69	1,263.
threat	0.04	0.49	0.07	69.
insult	0.47	0.91	0.62	1,170.
identity_hate	0.12	0.72	0.2	207.
micro avg	0.42	0.9	0.57	5,211.
macro avg	0.33	0.81	0.44	5,211.
weighted avg	0.49	0.9	0.63	5,211.
samples avg	0.05	0.08	0.06	5,211.

6 Conclusions

The BiLSTM model is optimized to have an high recall is performing good enough to make predictions for each label. Considering the low support for the threat label, the performance is not bad. See Table 2 and Figure 1: the threat label is only 0.27 % of the observations. The model has been optimized for recall because the cost of not identifying a injurious comment as such is higher than the cost of considering a clean comment as injurious.

Possibile improvements could be to increase the number of observations, expecially for the threat one. In general there are too many clean comments. This could be avoided doing an undersampling of the clean comment, which I explicitly avoided to check the performance on the BiLSTM with an imbalanced dataset, leveraging the class weight method.