Graph-Based Recommendation System

This project explores scalable and parallelized implementations of graph-based recommendation models: PinSage and PPNP.

How to Run

(TBD) On CUDA machine there is only like stupid 4GB of data so we use /scratch

Run the below to get the installations working

# Example for pip cache
mv ~/.cache/pip /scratch/$USER/pip_cache
ln -s /scratch/$USER/pip_cache ~/.cache/pip

# Example for .local (where pip installs user packages)
mv ~/.local /scratch/$USER/local
ln -s /scratch/$USER/local ~/.local

python main.py --epochs 120 --patience 75 \
               --data-root ./data/Planetoid \
               --log-csv run1.csv

Running the Code

Basic Usage

1.Without Benchmarking (Basic Training): python main.py --dataset Cora

2.With Benchmarking (Full Profiling): python main.py --dataset Cora --profile --track_memory --track_time

This will run the training with: Profiling enabled (--profile) Memory tracking (--track_memory) Detailed timing metrics (--track_time)

3.With Distributed Training (DDP):

# Without benchmarking
python main.py --dataset Cora --ddp

# With benchmarking
python main.py --dataset Cora --ddp --profile --track_memory --track_time

Additional useful options you can add: --epochs N: Set number of epochs (default is 200) --hidden N: Set hidden layer size (default is 64) --lr X: Set learning rate (default is 0.01) --dropout X: Set dropout rate (default is 0.5) --patience N: Set early stopping patience (default is 100)

For example, a full benchmarking run with custom parameters: python main.py --dataset Cora --epochs 100 --hidden 128 --lr 0.005 --profile --track_memory --track_time

The benchmarking output will include: A CSV file (appnp_training_log.csv) with detailed metrics A plot (training_plot.png) showing loss and validation accuracy Console output with: Per-epoch metrics Final benchmark summary Memory usage statistics Timing statistics

To document speedups, you can now: Run ./run_comparison.sh for a quick comparison Run python benchmark_parallel.py for comprehensive benchmarks Generate detailed reports with python benchmark_parallel.py --report