mogp-decomposition

The repository for the UAI 2021 publication 'Multi-output Gaussian Processes for Uncertainty-aware Recommender Systems'.

Souce code

• Multi-output / Multi-way GP
• Data loading functions

Basic usage

# Define the hyper parameters
hyper_params = {
    'I':I, 'J':J, 'K':K,      # the number of entities in each domain
    'emb_sizes': [8, 8],      # the embedding sizes
    'M': original=128,        # the number of inducing pairs
    'emb_reg': 1e-3,          # l2 norm on the embeddings
    'batch_size': 2**16,      # the size of training batches
    'obs_mean': Y_tr.mean(),  # the mean of target
    'lr': 1e-2                # the learning rate 
}  
gp_md = GPD(**hyper_params)
gp_md.save_path = './ml-1m_lrs/'+str(lr)+'_cv'+str(cv_id)+'/'
gp_md.build()
gp_md.train(X_tr, Y_tr, X_te, Y_te, n_iter=501)
gp_md.save()

Datasets

• ML-1M
• ML-10M
• Jester where each folder also contains the train/test split.

Experiments.

• ML-1M
  • Training script
  • Results summary
• ML-10M
  • Training script
  • Results summary
• ML-10M
  • Training script
  • Results summary

Note: github doesn't seem to be rendering the notebooks correctly. Consider downloading them for reviewing.