Missing Data 3

Handling Missing Data in Decision: A probabilistic approach

key ideas

  • A decision tree’s structure and notation

  • Review of imputation

    • Predictive value imputation

      • mean, median or mode

      • make assumption that features are indpendent

      • surrogate splits, partition data using another feature to

  • XG Boost

Expected Predictions:

  • impute all possible completions as once to avoid strong dist assumptions

  • consistent for MCAR and MAR

  • expensive, but density can help reduce

  • tractably compute the exact expected predictions

  • loss minimization

Experiments

  • for a single dataset, outperforms in general

Discussion

  • generally easier

  • given single dataset, of results, how much do we trust this?

  • what does this provide as an advantage

  • NP hard

How to miss data?: Reinformcent leanring for environgments iwith high obseration cost

Key points

Reinforcement learning

  • cost associated with making accurate observations

  • goal directed

  • RL agent tries to

Problem setting:

  • \(o \sim p_0(o_t |s_t; \beta)\)

  • beta is accuracy og obs

  • r is old reward

Scenario A:

  • observed cangle vs

Big picture: manipulating how the data collection

Discussion

  • survivorship bias?

  • right left imbalance for figure 3

  • simple pendulum example helped overcome the background lacking

  • figures

General

Try writing out a missingness graph for a problem of choice, some scenario where you imagine there would be missing data, or an example dataset that you can find.