This repository contains code for paper GLocal-K: Global and Local Kernels for Recommender Systems.

The proposed matrix completion framework based on global and local kernels, called GLocal-K, includes two stages: 1) pre-training an autoencoder using the local kernelised weight matrix, and 2) fine-tuning the pre-trained auto encoder with the rating matrix, produced by the global convolutional kernel. This repository provides the integrated implementation of two stages with two types of kernels on three benchmarks: ML-100K, ML-1M, and Douban.

Download this repository. As the code format is .ipynb, there are no settings but the Jupyter notebook with GPU.

• numpy
• scipy
• tensorflow (converted to version 1.x automatically in the main code)

1. Insert the path of a data directory on the main code by yourself (e.g., '/content/.../data').
2. Write down a dataset correctly among 'ML-1M', 'ML-100K', and 'Douban' on the main code.
3. There are no other things to do anymore, just try running the code and see it.

We evaluated our model GLocal-K on ML-100K, ML-1M and Douban with the metric RMSE, and the results are provided in our paper. In addition, we also tested our model using other metrics MAE and NDCG, which are widely used for rating prediction tasks.

• [ML-100K] - RMSE: 0.8889 / MAE: 0.6950 / NDCG: 0.9053
• [ML-1M] - RMSE: 0.8227 / MAE: 0.6421 / NDCG: 0.9288
• [Douban] - RMSE: 0.7208 / MAE: 0.5622 / NDCG: 0.9435

1. Harper, F. M., & Konstan, J. A. (2015). The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis), 5(4), 1-19.
2. Monti, F., Bronstein, M. M., & Bresson, X. (2017, December). Geometric matrix completion with recurrent multi-graph neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems (pp. 3700-3710).