Practical Question on Bayesian Ensembles

Anyone experienced in combining the outcome of several weak models into one? I am looking for practical tips that could help.

  • several smaller models are often easier and faster to build, test and maintain
  • sometimes the business experts do have different opinions / models they belief in
  • smaller models need less data to train and converge
  • there is a potential to reduce overfitting

The most naive way I can think of is a weighted score over several models. A weighted mixture would be ok too I assume.

But, will it work? Has someone experienced with real life problems ?


Have you checked out the notebook on Bayesian model averaging? That might be a decent place to start.

Thanks @cluhmann - that’s a very interesting example. Gonna work through it the next days :slight_smile:

1 Like