Project idea from Awesome Deep Learning Project Ideas
The training data set was originally downloaded from "StackLite: Stack Overflow questions and tags" and is formally licenced by Stack Exchange, Inc. under cc-by-sa 3.0. It contains the question score and answer count as well as the anonymous ID of its owner. The neural net tries to map this vector to one of the 50 frequently used question tags like java, c++ or html.
The neural net was implemented as computational graph with the popular machine learning library TensorFlow. You can find my model in the following Python module:
src/model.py. The below picture shows the network architecture. It consists of four hidden layers with 10, 12, 24 and 48 neurons, where each neuron
has a ReLU activation. Further neural nets output layer holds one neuron
for each question tag and applies Softmax function to their activation for classification purpose.
To train model, simply run the following command in the root folder of this project. Therefore Python 3 is recommended and Googles TensorFlow and matplotlib are required.
$ python src/model.pyThis models reaches an accuracy of over 85% for the train and test data set after 20000 iterations of training. The below picture shows models loss in relation to its training epochs. Under data/trained_models you can find this neural net as pre trained model with its adjusted weights and biases. Use TensorFlows tf.train.Saver to load this model and make your own predictions against this tagged questions.
