Comprehensive Review of Multiclass Text Classification using the 20 Newsgroup Dataset

This dissertation showcases a comprehensive study of machine learning and deep learning algorithms on multiclass text classification using the 20Newsgroup dataset. The CRISP-DM methodology was followed, and a detailed step-by-step approach was taken to document every step from data collection to result evaluation. EDA was carried out before and after cleaning and preprocessing and the result was visualised in a word cloud to see the tokenized text. The dataset was pre-processed using tokenization, removal of stopwords, changing to lowercase and vectorization using TF-IDF. The models implemented are Naïve Bayes, KNN, XG Boost, Logistics Regression, Random Forest, Decision Tree, SVM, CNN, ANN, RNN and LSTM. It is interesting to note that SVM had the best accuracy and performed better than any of the Deep Learning models. The confusion matrices provided more details as to where each of the models struggled. Further investigation can be done on this work to find out why the deep learning models did not outperform the machine learning models. This work presents an accurate comparative analysis which can be validated by running the code attached.