What it does
Using an advanced Long Short Term Memory Network, we analyze the raw headline text for a news story to determine if it is real or fake news.
How we built it
Using the Tensorflow and Keras libraries, we have a character level neural network with 3 layers of Bidirectional Long Short Term Memory cells, along with 2 Hidden Densely Connected Layers. Dropout and Gaussian Noise Layers help to avoid over fitting. The network was then trained on ~20,000 headlines samples (50% real, 50% fake). 20% of these samples were withheld from training and used to validate the network.
Challenges we ran into
Network design and tuning Finding an appropriate and accurate dataset Convincing people we built this during the hackathon.
Accomplishments that we're proud of
Has an accuracy of 92% on the validation set. Passing many tests from data acquired from outside of the 2 datasets from which the training samples came from
What we learned
There exists sufficient data within the raw headline text of a story to determine if it is real or fake news Long Short Term Memory Networks can be useful in fake news detection
What's next for Fighting Back Against Fake News
Analyzing text snippets of a news article Retraining on additional data for a more accurate network Rolling out to journalists and fact checkers
Replace the string at the end of the link with the headline to analyze. Number returned is the computed percent chance that the headline is fake news.