NBI - Network-Based Inference Recommendation

A C implementation of the Network-Based Inference (NBI) algorithm for personalized recommendations using bipartite network projection.

Algorithm

Based on the paper:

Tao Zhou, Jie Ren, Matúš Medo, Yi-Cheng Zhang, "Bipartite network projection and personal recommendation", Phys. Rev. E 76, 046115 (2007).

The NBI algorithm performs personalized recommendations through a two-step resource diffusion process on a user-item bipartite network:

1. Resource Allocation: Assign initial resource (f=1) to items selected by the target user
2. Items → Users Diffusion: f(user) = Σ f(item) / k(item)
3. Users → Items Diffusion: f'(item) = Σ f(user) / k(user)
4. Ranking: Sort items by final score, excluding already selected items

Build & Run

# Compile
gcc nbi.c -o nbi

# Run with N recommendations per user
./nbi <N>

# Example: generate top-10 recommendations
./nbi 10

Data Format

Input files use tab-separated format:

UserID    ItemID    Timestamp

File	Description
training.txt	Training data for building the network
check.txt	Test data for evaluation
result.txt	Output recommendations (generated)

Dataset

The included dataset is from MovieLens, randomly split into training and test sets.

Output

Recommendations: 10, Hits: 11044, Precision: 0.1829

• Recommendations: Number of items recommended per user
• Hits: Total correct predictions in test set
• Precision: Hit rate = Hits / (N × Users)

References

• Original Paper (Phys. Rev. E)
• arXiv Preprint
• MovieLens Dataset