Topic-Modelling GCN + LDA

This repository explores Latent Dirichlet Allocation methods for text based classification employing various Graph Convolutional Networks.

Dataset

Download from: https://drive.google.com/file/d/10kx3z3bjYFoeRjjg1_DZOAP39Jln0BCh/view?usp=sharing and keep under TestSGC/ after extracting.

Step:: 1 Pre-Processing

For pre-processing and arranging the dataset into DataFrame except for 20ng and ohsumed (which are done as given in the code step_1_data_to_pandas_normal.py) remaining datasets have their iPyNB in their respective dataset directories.

Step:: 2 LDA Feature Vector

step_2_topic_modelling.py

Step:: 3 Gathering LDA Feature Vector into a Composite Feature Matrix

For matching the feature matrix of GCN in "Graph Convolutional Networks for Text Classification's" implementation, we have used their file used for indicating document names, training/test split, document labels. Each line is for a document. These files are stored under document_information.

Features

Our approach is an ensemble work of features from LDA and Graph Convolution.

Code Description

• LDA features obtained from topic_modelling.py
• The data is then converted to LDA probability matrix using data_to_pandas.py
• The GCN features are calculated considering various parameters and stored for training.
• SGC/downstream/TextSGC/ contains the model files for training the different network architectures.
• The model also involves the use of skip-architecture which improves model performance.

The features obtained from LDA together with the GCN features are merged in a systematic fashion to obtain a feature rich map which is then fed into a custom-build model. Experiments were carried on to obtain optimal results.