Heterogeneous Multi-output Gaussian Processes

This repository contains the implementation of our Heterogeneous Multi-output Gaussian Process (HetMOGP) model. The entire code is written in Python and connected with the GPy package, specially useful for Gaussian processes. Our code consists of two main blocks:

• hetmogp: This block contains all model definitions, inference, and important utilities.
• likelihoods: General library of probability distributions for the heterogeneous likelihood construction.

Our tool is a novel extension of multi-output Gaussian processes for handling heterogeneous outputs (from different statistical data-types). The following distributions are already available to be used: [Gaussian, Bernoulli, Heteroscedastic Gaussian, Categorical, Exponential, Gamma, Beta, Poisson]. We expect to release code for Student, Ordinal, Geometric, Binomial, Multinomial, Truncated Gaussian, Wishart and Dirichlet likelihood distributions as soon as possible. If you want to contribute or include a new likelihood function, please follow the instructions given below to add your new script to the likelihoods module.

Please, if you use this code, cite the following paper:

@inproceedings{MorenoArtesAlvarez18,
  title =  {Heterogeneous Multi-output {G}aussian Process Prediction},
  author =   {Moreno-Mu\~noz, Pablo and Art\'es-Rodr\'iguez, Antonio and \'Alvarez, Mauricio A},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS) 31},
  year =   {2018}
}

Usage

Our Python implementation follows a straightforward sintaxis where you only have to define a list of input and output values, build the heterogeneous likelihood with the desired distributions that you want to predict and call directly to the model class. That is

• Output and input data definition:

Y = [Y_real, Y_binary, Y_categorical]
X = [X_real, X_binary, X_categorical]

• Heterogeneous Likelihood definition:

likelihood_list = [HetGaussian(), Bernoulli(), Categorical(K=3)]

• Model and definition:

model = HetMOGP(X=X, Y=Y, Z=Z, kern_list=kern_list, likelihood=likelihood, Y_metadata=Y_metadata)

A complete example of our model usage can be found in this repository at notebooks > demo

New Likelihoods

The heterogeneous likehood structure (based on Eero Siivola's GPy release and GPstuff) permits to handle mixed likelihoods with different statistical data types in a very natural way. The idea behind this structure is that any user can add his own distributions easily by following a series of recommendations:

1. Place your new_distribution.py under the likelihood directory.
2. Define the logpdf, first order dlogp_df and second order derivatives d2logp_df2 of your log-likelihood function.
3. Use var_exp and var_exp_derivatives for approximating integrals with Gauss-Hermite quadratures.
4. Code your predictive and get_metadata methods to have all available utilities.

Examples

• Missing Gap Prediction: Predicting in classification problems with information obtained from parallel regression tasks.

• London House Prices: Spatial modeling with heterogeneous samples. This is a demographic example where we mix discrete data (type of house) with real observations (log-price of house sale contracts).

Potential Applications

We have collected many ideas about possible applications of our heterogeneous multi-output GP model.

Contributors

Pablo Moreno-Muñoz, Antonio Artés-Rodríguez and Mauricio A. Álvarez

For further information or contact:

[email protected]