Missing data is a major problem for anyone working with large datasets. There have been many approaches to dealing with missing data, from list-wise deletion of missing data to imputation via multivariate logistic regression. This packages provides a framework to use Bayesian networks to predict, evaluate and impute missing or held-out data using network-based expectation maximization. This method is particularly well suited for highly dimensional categorical data that doesn't behave well with traditional imputation approaches.
In the following vignette, I will walk through an example in which we:
- Create the structure of a bayesian network
- Hold out 25% of the values of the
jobfeature in the dataset, creating missing data to predict. - Train the bayesian network with the truncated dataset
- Predict the "missing" values using Most Probable Explanation technique
- Evaluate the model
- Visualize the evaluation process
- Impute the data
See the example in the vignettes folder for further instruction and information