Kernel methods are a fairly mature branch of machine learning, but are heavily used in e.g. bioinformatics or time series modelling. Our latest contribution describes how learning with multiple kernel functions (representations) can be implemented in an efficient way. The latest Neurocomputing issue features our paper Approximate multiple kernel learning with least-angle regression in open access, describing the design and various applications of the Mklaren algorithm. Highlights:

  • Efficient kernel selection in linear regression,
  • competitive performance among kernel approximation methods,
  • blueprint for applications in time series modelling, text mining and biological sequence regression,
  • perhaps most importantly, comprehensive Mklaren Python library. It implements several kernel approximation and multiple kernel learning methods in a memory-efficient way.

Big thanks to my advisor and colleague dr. Tomaž Curk for constructive collaboration as this work developed over time!