Fake-News-Detector

Test out the model at this link.

Learning basics of Deep Learning with TF-IDF and Fake News Detection. Original project at https://data-flair.training/blogs/advanced-python-project-detecting-fake-news/, but will do some fun twists.

Goals

Overall: The goal is to have different types of models that will predict whether a post, article, or other text is considered fake news or not. I will try to split each article into a different topic (currently within political atmospheres, will hopefully spread to non-political news as well). Goal is to predict with >95% accuracy, and reduce the number of false negatives that exist (aka. classifying text as real when it is actually fake).

1. Create a Fake News classifier by implementing basic NLP preprocessing techniques (such as vectorization) and training on an SGD-augmented linear model. Should be applicable to any news article. Slight performance tuning adjustments to prioritize FPs/FNs may need to be done. - in progress in model.ipynb
2. Create a web scraping pipeline where if someone puts in an article link, then it will extract the article and print out the inference
3. Create a new model that focuses on short-form videos and/or small statements
4. Create models that focus on specific topics (like war, economy, disasters, etc.)

The goal is to have different types of models that will predict whether a post, article, or other text is considered fake news or not. I will try to split each article into a different topic (currently within political atmospheres, will hopefully spread to non-political news as well). Goal is to predict with >95% accuracy, and reduce the number of false negatives that exist (aka. classifying text as real when it is actually fake).

Progress

Got two models running and deployed on a simple website (see above).

The first is for long articles, using a SGD-boosted Logistic Regression classifier to determine whether text is fake or real. This model works best on longer bodies of text as compared to shorter paragraphs, and has around 96% accuracy.

The second is for Twitter posts, specifically shorter-form posts and comments over the site. This model uses a Histogram-based Gradient Boosting Classifier, but it performs worse than the long-text model (~80% accuracy) due to the shorter content. Some features may need to be removed due to web scrape parsing issues, so still up for some change.

Next: Want to add transcripts of videos from various social media platforms, analysis of pictures, and support for Instagram, Tiktok, Facebook, etc.

Requirements: requirements.txt

Data Citation: in data_sources.md

Also check out another project of mine, curly-waffle-spark