Project Type: Individual 

Project Duration:  1 month

Software Used: Py Charm

Languages Used: Python 3.7

Advanced Machine Learning application for text categorization using Deep Neural Networks and giving a detailed analysis on the pre-existing DNN’s, comparative analysis of the custom made DNN with the pre trained DNNs.

The programming language used to develop the DNN model is python 3.7 with stream lit library for GUI implementation. The other Machine Learning libraries used to build the model are TensorFlow and K eras. The interface used is Py Charm for python programming with Anaconda as virtual environment.

Deep Neural Network was performed on the famous BBC News data set using python, TensorFlow, K eras and stream lit library. The data set has 5 classes and 2225 sample observations. The data is in text format and does not need any pre-processing as the K eras has already pre-processing methods for the data to use it readily. Then tokenization was used to convert each word into numerical array format and encode the unique classes into numerical format as 0,1,2,3,4 for the 5 classes using Label Encoder and .to categorical() method of K eras. We then split the data into 80% training and 20% testing and create the DNN model. After various trial and errors, the optimum unit is between 500-550 for 32 batch size.

Two pre trained models were used – Gnews and NNLM, where after training the Gnews model again on the BBC data set gives the better prediction and classification performance. However, NNLM performance doesn’t change and still gives the worst performance.

Out of the three models - 2 pre trained and 1 custom built, the later gives the best performance than the pre-trained models and can be used to classify the news precisely. The custom DNN model is the best model.