Experimental Comparison of Recommender Systems

Project summary

The goal of the project is to evaluate several state-of-the-art recommender systems applied to benchmark datasets. Particular attention will be given to convolutional neural networks on graphs.

This is a Final Year Project by See Jie Xun, supervised by A/P Xavier Bresson. The project website can be found here.

Packages required (Python 3)

• tensorflow >= 1.13 (If possible, use tensorflow-GPU for faster training of models)

• numpy >= 1.14

• scipy >= 1.1

• scikit-learn >= 0.20

• pandas >= 0.23

• h5py >= 2.7

Data

The datasets are already contained in the repository, under the "gcmc_adaptation/data" folder. The Douban, Flixster, and YahooMusic datasets are taken from the repository for the Graph Convolutional Matrix Completion method by van den Berg et al.

The Facebook dataset needs to be processed to generate an adjacency matrix and user features, for us to be able to train a GCN on it. To do this, run the gcmc_adaptation/data/facebook/processing.py script, which will save the adjacency matrix as adjacency2.npy in the same folder. The user features have already been generated and are saved as user_features.npy.

Scripts to use

To run a single experiment, use the script gcmc_adaptation/train.py. Hyperparameter and experimental settings need to be specified at the command line when running the script. For example, use the command to run an experiment on the Facebook dataset:

python train.py -d facebook --accum stackRGGCN --dropout 0.5 -e 1000 --hidden 50 25 --num_layers 2 --features --testing

Note: The --accum flag selects the GCN architecture to be used to generate representations for each node based on the graph structure and node features. The primary focus of this project was on investigating the RGGCN architecture as applied to this graph autoencoder framework for recommender systems. This can be selected by specifying --accum stackRGGCN.

To evaluate the results of a set of hyperparameters, averaged over a number of iterations, use the script gcmc_adaptation/evaluate_hyperparameters.py. The hyperparameters and number of iterations can be edited within this script.