What is it about?
Many important machine learning problems can be phrased as pairwise learning, i.e. modeling properties of two objects. For example, predicting whether a person will like a movie, whether a small chemical compound will bind a protein or if a certain species of bee can pollinate a type of flower. Pairwise kernels can be used to combine the feature representations of the two objects into a joint feature representation of the pair of objects. Using this representation, simple, though powerful learning methods can be constructed.
Featured Image
Why is it important?
We provide a theoretical foundation on how pairwise learning methods are related, showing that Kronecker kernel ridge regression, two-step kernel ridge regression and a simple linear filter are just instances of kernel ridge regression. These methods can be used for a variety of machine learning tasks: network inference, collaborative filtering, multi-task learning, zero-shot learning etc.
Perspectives
In this article we have tried to construct a theoretical framework for the Kronecker kernel-based framework our research groups collaborated on for many years. The paper also reviews these methods, with an emphasis on efficient implementations. We hope that this paper is a useful reference for people who want to apply pairwise learning in new domains or as a basis to build new theory.
Michiel Stock
Ghent University
Read the Original
This page is a summary of: A Comparative Study of Pairwise Learning Methods Based on Kernel Ridge Regression, Neural Computation, August 2018, The MIT Press,
DOI: 10.1162/neco_a_01096.
You can read the full text:
Contributors
The following have contributed to this page
