📖 Documentation for project: Detection of fake Ukrainian news

📌 Description:

This is a student project for the Machine Learning course, aimed at developing an application for detecting fake news.

🎯 The main goal:

Create a machine learning model capable of analyzing Ukrainian news texts and determining their authenticity.

✅Approach

1️⃣ Data Collection

• Searching for available Ukrainian news datasets in internet
• Downloading datasets (if we find it)
• Analyze the data structure and identify key fields (URL, title, text, source, date). What sources the news is from, etc.

2️⃣ Data Preprocessing

• Extracting relevant information for further work. For example:
  • Title of news
  • Text of news
  • Time when the news was published
  • Additional info that will be needed later
• Removing duplicates and unnecessary data
• Analyzing data distribution and identifying news sources. We divide them into the following
  • Sources we trust:
    • TSN
    • BBC
    • Суспільне
    • ҐРУНТ
    • STERNENKO
    • Лачен пише
  • Sources we do not trust:
    • Вокс Україна
    • Война с фейками
• Balancing classes (if necessary)

3️⃣ Model Development

• Selection of a machine learning algorithm by the selection method.
• Training of the selected algorithm
• Determining its performance.
• Optimization of hyperparameters to increase the accuracy.
• (Optional) If the approach fails, consider:
  • Using an LLM model.
  • Falling back to GPT-based classification.

4️⃣ Creating a WebUI for easier interaction

• Using streamlit for this purpose.

📊 Datasets Used

1️⃣ Ukrainian Fake and True News

• Source: Kaggle
• Description: Contains Fake and True news about Russo-Ukrainian war

2️⃣ Ukrainian News

• Source: Hugging Face
• Description: A dataset of news articles downloaded from various Ukrainian websites and Telegram channels

• Divided the Dataset from Hugging Face

  • Source:

    • Filemail. A modified version of the Hugging Face dataset, divided into 22 parts, each containing 1 million rows.
    • Filemail. A modified version of the Hugging Face dataset, divided into 46 parts, each containing 500 thousand rows.

  • Using this code for that purpose.

🚀 Development Journey

⭐ 2025-02-03 – Team Formation.

👥 Team Members:

1. PM: voinskyi
2. Data engineer: Yul4onok
3. Data scientist: highbrow-228

💡2025-02-06 – Development of the Idea to Tackle Disinformation.

🔍2025-02-14 – Searched for datasets with Ukrainian news, but failed.

📰2025-02-16 – The first try to scrap news from TSN.ua

• News is scrapping very slowly so we decided to find other ways of collecting data.

🤔2025-02-17

• Came across the problem of labeling data (how to label a huge amount of data (fake or real) by ourselves?).
• The second try to find the dataset and already successful. (We found ukrainian-news on HuggingFace and Ukrainian News on Kaggle).
• We consider the idea of labeling data this way: Choose sources we trust and mark that news as "real", then choose some suspicious sources and mark them as "fake".

🗂️2025-02-18

• Trying to download ukrainian-news dataset on HuggingFace but came across of lacking resources (the dataset is really large (22 milion rows) and takes a lot of RAM to process). So we decide to divide the dataset into 23 subsets for better processing.

📌 Next Steps... Stay Tuned!