app.py

import gradio as gr
import pandas as pd
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from textblob import TextBlob
import os
from huggingface_hub import login

# Get the Hugging Face API token from the environment variable
hf_token = os.getenv("pasavectoi")
login(hf_token)

# Load the dataset from the local file
data = pd.read_csv('twitter_dataset.csv').head(1000)

# Calculate sentiment polarity and popularity
data['Sentiment'] = data['Text'].apply(lambda x: TextBlob(x).sentiment.polarity)
data['Popularity'] = data['Retweets'] + data['Likes']
data['Popularity'] = (data['Popularity'] - data['Popularity'].mean()) / data['Popularity'].std()
data['Popularity'] = data['Popularity'] / data['Popularity'].abs().max()

# Load the fake news classification model
model_name = "hamzab/roberta-fake-news-classification"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Process tweets in batches to avoid memory issues
batch_size = 100
predictions = []
for i in range(0, len(data), batch_size):
    batch = data['Text'][i:i + batch_size].tolist()
    inputs = tokenizer(batch, return_tensors="pt", padding=True, truncation=True, max_length=128)
    inputs = {key: val.to(device) for key, val in inputs.items()}
    with torch.no_grad():
        outputs = model(**inputs)
    predictions.extend(outputs.logits.argmax(dim=1).cpu().numpy())

data['Fake_News_Prediction'] = predictions
data['Credibility'] = data['Fake_News_Prediction'].apply(lambda x: 1 if x == 1 else -1)

# Define the prediction and recommendation function
def predict_and_recommend(title, text, visibility_weight, sentiment_weight, popularity_weight):
    # Adjust weights and calculate the final score
    total_weight = visibility_weight + sentiment_weight + popularity_weight
    visibility_weight /= total_weight
    sentiment_weight /= total_weight
    popularity_weight /= total_weight

    # Update final visibility score with user-defined weights
    data['User_Final_Visibility_Score'] = (
        data['Credibility'] * visibility_weight +
        data['Sentiment'] * sentiment_weight +
        data['Popularity'] * popularity_weight
    )
    # Sort and randomly sample 10 recommendations
    top_100_data = data.nlargest(100, 'User_Final_Visibility_Score')
    recommended_data = top_100_data.sample(10)

    return recommended_data[['Text', 'User_Final_Visibility_Score']]

# Set up Gradio interface
iface = gr.Interface(
    fn=predict_and_recommend,
    inputs=[
        gr.Textbox(label="Title"),
        gr.Textbox(label="Text", lines=10),
        gr.Slider(0, 1, 0.5, label="Visibility Weight"),
        gr.Slider(0, 1, 0.3, label="Sentiment Weight"),
        gr.Slider(0, 1, 0.2, label="Popularity Weight")
    ],
    outputs="dataframe",
    title="Customizable Fake News Recommendation System",
    description="Adjust weights to receive customized tweet recommendations based on visibility, sentiment, and popularity."
)

iface.launch()