The context is this - I'm considering doing a part time PhD in statistical learning and today I've met up with a prospective supervisor who suggested that I think about causality in machine learning as (in his words) a lot of people are doing deep learning/neural network stuff and less so in time series learning and causality inference. Causality inference is less developed in machine learning and it's an area I could contribute to.

So I've spent some time googling and thinking and I came away with these elements that I would ideally like to see in a learning algorithm.

  • Can handle multi-dimensional time series data
  • Online learning
  • Time varying relationships
  • Can use regularisation
  • Allows causality inference (actually, this may overlap with regularisation in some way to prevent overfitting)

This to me sounded like a state-space/dynamic model of some sort. As an example, I could extend the Kalman filter in some way to include the above. Any idea on more specific algorithms I should have a look at?


If you're thinking about extending the Kalman filter you may be interested in these publications as a starting point. They check most of the boxes except the causality, but that is where you come in:

  1. Särkkä, S. (2013). Bayesian filtering and smoothing (Vol. 3). Cambridge University Press. Online version.

  2. Van Vaerenbergh, S., Lázaro-Gredilla, M., & Santamaría, I. (2012). Kernel recursive least-squares tracker for time-varying regression. IEEE transactions on neural networks and learning systems, 23(8), 1313-1326. Online version.

[Disclaimer: I'm author of the second one.]


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.