e-vdb / toxic-comment-classification

Multi-label classification on toxic-comment dataset

toxic-comment-classification

Summary

Multi-label classification of Wikipedia comments. The purpose is to predict the toxicity of these comments among 6 categories :

• toxic
• severe_toxic
• obscene
• threat
• insult
• identity_hate

Dataset

Dataset available on https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data

Repository content

• toxicCommentClassification.ipynb : Jupyter notebook
• ML_model : folder of Python scripts
  • model.py : saves a Linear SVC model trained on the dataset with Pickle (finalized_model.sav)
  • f1_score_test.py : loads trained model and evaluates the f1_score on the test dataset
  • dev_tests.py : loads trained model and gives predictions of toxicity
  • user.py : loads trained model and gives predictions of toxicity for a console entry

Jupyter notebook

• Exploratory data analysis
• Data preprocessing
• Natural Language Processing
  • Standard CountVectorizer
  • Tfidf Vectorizer
• Multi-label classifier using a MultiOutputClassifier from sklearn.multioutput
• Machine learning algorithms
  • Multinomial Naive Bayes Classifier
  • Linear Support Vector Machine Classifier
  • Logistic Regression
• Evaluation of the f1 score (weighted)
• Parameters tuning with GridSearch CV

TASK LIST

• Implement machine learning algorithm using Scikit-learn
• Implement deep learning algorithm using Keras
• Deploy model