Add strategies to account for missing data

[biprateep]

Here are a few promising methods:

• https://arxiv.org/pdf/1805.07405
• https://proceedings.neurips.cc/paper_files/paper/2018/file/411ae1bf081d1674ca6091f8c59a266f-Paper.pdf
• https://arxiv.org/pdf/2206.01640
• https://proceedings.neurips.cc/paper/1993/file/a8ecbabae151abacba7dbde04f761c37-Paper.pdf
• https://openreview.net/pdf?id=J7b4BCtDm4
• https://en.wikipedia.org/wiki/Kalman_filter
• https://link.springer.com/chapter/10.1007/978-3-030-63836-8_19
• https://apps.dtic.mil/sti/tr/pdf/ADA295618.pdf

[biprateep]

Another strategy will be to train on curated data but during inference, replace the missing inputs by values drawn drawn from the feature space and then marginalizing the results. PZFlow takes this approach (see: https://github.com/jfcrenshaw/pzflow/blob/main/docs/tutorials/marginalization.ipynb). Need to look into if this is theoretically a sound thing to do.