This project is about news headers classification (text mining field), but the project core was preprocessing and feature engineering, you can expect:

• target variable analysis
• a lot of data preprocessing (cleaning with regex, stoplist building, lemmatization, all verified with word clouds)
• a lot of feature engineering
  • titles length measures (before and after preprocessing, mean number of words, normalized)
  • sentiment estimated with VADER (normalized)
  • named entity recognition with SpaCy (groups of tags, most frequent tags)
  • n-grams (most frequent unigrams and bigrams)
• GloVe word embeddings (100-dimensional vectors)
• model validation and architekture choosing
• final two input neural network architecture:
  • first branch based on pretrained GloVe embeddings, 1D Convolutions and 1D Max pooling
  • second branch based on engineered features (65 in total, sparse) and dense layer
  • concatenation to main branch with two more dense layers with regularization and dropout
  • softmax classifier with 4 outputs
• model evaluation and amazing test set results (at least 90% accuracy for each class, test set confusion matrix below)
• Jupyter Notebook in English with ton of python code :)

• Neural Networks
• NLP
• notebook
• Python