Facebook Friend Recommendation System

This project uses machine learning techniques to predict missing links in a social graph, specifically for recommending friends on Facebook. The task involves link prediction on a directed graph using features derived from social interactions.

Dataset

• Source: Facebook's recruiting challenge dataset on Kaggle.
• Columns:
  • source_node: The user sending the friend request.
  • destination_node: The user receiving the friend request.

Problem Statement

• Objective: Predict missing links (recommend users) in a directed social graph, mapping it to a supervised learning problem.

Key Techniques

1. Feature Engineering:

   • Generated features like the number of followers, mutual followers, PageRank, Katz score, Adamic/Adar index, and more.
   • Applied Singular Value Decomposition (SVD) on the adjacency matrix for dimensionality reduction.

2. Graph Algorithms:

   • Jaccard Index and Cosine Similarity to measure similarity between users.
   • Shortest Path to estimate the connection distance between users.
   • PageRank, HITS Algorithm, and Katz Centrality to compute node importance.

3. Supervised Learning:

   • Trained models using Random Forest Classifier and XGBoost on the extracted features.
   • Evaluated performance using F1 score and Confusion Matrix.

Results

• Best Model: Achieved high F1 score and accuracy using Random Forest and XGBoost classifiers.
• Evaluation: Precision and recall metrics indicated strong performance in recommending friends based on graph features.

Conclusion

The friend recommendation system effectively predicts missing links in a social graph by leveraging advanced graph-based features and machine learning algorithms. The system can recommend the most likely friends based on their probability of connection.