Multitask learning in GPFlow with missing inputs?

Question

Is it possible to do multitask learning with GPFlow where some inputs are missing? Specifically, I am trying to fit spatial data from several related individuals, but the data are not at identical input locations for all individuals. I know I ought to be doing hierarchical GPs here, but they tend not to scale well. I was hoping that multitask learning might be used instead, though the use case does not map exactly onto the typical application of this method.

Answer 1

There is currently not a model in GPflow that does this. However, GPflow does provide tools to easily implement this. Two suggestions:

• Use a multioutput kernel, and for missing datapoints you set your observation variance to infinity.
• Define a multioutput kernel, and specify a custom Kuf for which the requested outputs are passed together with the corresponding input.