This project builds multiple machine learning models to detect fake news articles and also includes a news summarizer.
The system is trained on a dataset of over 20,000 fake and real news articles scraped from various sources. Several models are trained and evaluated including Logistic Regression, Decision Trees, Random Forest, Gradient Boosting, and SVM.
In addition to fake news classification, the project also summarizes news articles by extracting key sentences.
The main datasets used are:
Fake.csv- Contains over 20K fake news articlesTrue.csv- Contains over 20K real/true news articles
The data_processing.py module handles loading these datasets, combining them, and adding additional real and fake samples using NLP techniques:
- Uses the Google News API to fetch the latest news articles on trending topics
- Scrapes the article content from the website
- Runs an NLP summary algorithm to get the key sentences
- Saves summaries to
summarys.csv
- Defines a large corpus of text related to technology, ethics, environment etc.
- Builds a next-word dictionary from this corpus
- Uses the dictionary to generate random fake paragraphs based on a seed topic word
- Saves fake summaries to
random_paragraphs.csv
The final output is a large CSV file combined.csv containing real and fake news samples.
The main preprocessing steps are:
- Load fake and real news CSVs
- Add fake/real labels
- Split off small sample for manual testing
- Concatenate into one dataset
- Shuffle rows
- Use Google News API to get latest articles
- Scrape article content
- Run NLP summarization algorithm
- Save summaries to
summarys.csv
- Build a next-word dictionary from a text corpus
- Use dictionary to generate random fake paragraphs
- Save to
random_paragraphs.csv
- Concatenate real summaries, fake summaries, and loaded dataset
- Reshuffle all rows randomly
- Clean text by removing special characters, punctuation etc.
- Convert text to TF-IDF vectors
The project trains an ensemble of different ML models on the fake news classification task, including:
- Logistic Regression
- Decision Tree
- Random Forest
- Gradient Boosting
- SVM with Linear Kernel
Using an ensemble approach provides several key advantages:
By combining multiple models together, the overall accuracy is improved compared to any individual model. Each model can capture different nuances in the data.
Every model has its own limitations. For example, decision trees can overfit easily. By combining with other models like random forests and boosting algorithms which reduce overfitting, we can overcome these individual weaknesses.
Ensembling multiple models reduces variance and makes the overall predictions more robust. For example, if a new data point causes one model to fail, the other models can compensate and still provide the right prediction.
The models are combined using simple majority voting - if 3 or more models predict a news article as Fake, we classify it as such. The threshold can be adjusted to tradeoff between precision and recall.
By training an ensemble of diverse ML models instead of relying on any single technique, our fake news classifier is more accurate, robust and resilient to new/unseen data. The ultimate goal is to maximize detection rates so harmful fake content can be effectively moderated.
The models are serialized to .sav files using Pickle to not train the model every single time we run the program.
The main entry point is main.py. Usage details are:
python main.py
A manual testing loop is also included to get predictions on custom user-inputted news articles.