ReviewRadar

ReviewRadar: Fake vs Genuine Review Detection

ReviewRadar is a web application that allows users to send links to review pages, and classifies each review as real or fake based on a user-defined strictness level. The application uses crawl4ai to scrape reviews from the provided URLs, processes them through an OpenAI LLM extraction strategy to generate the required JSON format, and then passes the reviews to a pretrained classifier with 88% accuracy for classification.

Users can also input individual reviews for classification directly on the app. Additionally, the app provides an open API that allows developers to use the model for review classification through a simple JSON interface.

Key Differences Between Fake and Real Reviews:

1. Overly Positive Language (Fake): Exaggerated praise with phrases like "Love this!", "Amazing!", and "10 Stars", often without specifics.
2. Repetitive Phrasing (Fake): Common phrases like "I love" and "the only problem is" repeated across reviews.
3. Personal Pronouns (Fake): Frequent use of "I", "we", or "my" to create a forced personal connection.
4. Contradictory/Incomplete Sentences (Fake): Some reviews seem cut off or make contradictory statements.
5. Time Mentions (Fake): Fake reviews often mention usage duration to imply long-term experience.
6. Specific & Critical Feedback (Real): Real reviews focus on specific product features, highlighting both pros and cons.
7. Practicality Focus (Real): Emphasis on product functionality and day-to-day use.
8. Balanced Opinions (Real): Mix of positive and negative points.
9. Short & Objective (Real): Concise and direct feedback, avoiding unnecessary embellishment.

Features

• Review Page Analysis: Users can submit URLs containing reviews, which are scraped and classified as real or fake based on user-selected strictness levels.
• Individual Review Testing: Users can input single reviews and get immediate feedback on their authenticity.
• Data Visualization: The app offers insightful visualizations, such as pie charts and histograms, to help users analyze the distribution of real and fake reviews on a page.
• Open API: Developers can access an API endpoint to integrate the model into their own applications for fake review detection.

API Endpoint

POST /api/openapi-verify-review

Developers can use this endpoint to verify reviews through a simple JSON interface. The input includes the review text and a threshold level (high, medium, or low), and the output is a classification of the review as real or fake.

Example Request:

{
  "review": "Great product, highly recommend!",
  "threshold": "high"
}

Example Response:

{
  "analyzed_reviews": {
    "review_text": "Great product, highly recommend!",
    "is_fake": false,
    "confidence": 0.92
  }
}

Model Approach

Objective: To classify reviews as fake or genuine using text-based features.

Algorithm:

• Support Vector Classifier (SVC), chosen for its effectiveness in text classification tasks.

Process

1. Data Preparation:

   • Loaded a dataset of reviews and preprocessed the text (removing punctuation, filtering stop words).
   • Split the dataset into training and testing sets.

2. Text Processing:

   • Used a text processing function to prepare reviews for vectorization.
   • Applied CountVectorizer to convert the text into a bag-of-words model.

3. Model Training:

   • Created a pipeline combining:
     • CountVectorizer: Converts text into numerical vectors.
     • TF-IDF Transformer: Scales vectors by term frequency-inverse document frequency.
     • SVC Classifier: Trained on the vectorized data.
   • Trained the model on the preprocessed data.

4. Model Saving:

   • Saved the trained model using joblib for later predictions.

Outcome:

• The model achieves 88% accuracy in classifying reviews as real or fake.
• Successfully loads the model for inference, allowing users to analyze new reviews.

How It Works

1. User Submission: Users submit a URL containing reviews or an individual review.
2. Scraping and Parsing: The application scrapes reviews from the URL using crawl4ai, and the scraped content is processed through an OpenAI LLM to generate structured JSON data.
3. Review Classification: The structured data is passed through the pretrained SVC classifier to identify fake reviews.
4. Visualization: The app displays the results, offering visual insights through pie charts and histograms.

Usage Example

Users can submit a link to a review page and get a visual breakdown of real vs fake reviews in a pie chart and histogram format, based on their chosen strictness level (high, medium, low). They can also manually input a review to check its authenticity.

Conclusion

ReviewRadar simplifies the process of detecting fake reviews, offering both a web interface and a developer-friendly API. Its robust classification pipeline, powered by SVC, ensures high accuracy and reliable results, making it an essential tool for consumers and developers alike.