A missing value occurs when one or more than one features are missing for some of the instances in your input data. This may happen because, for example, a user forgot to fill in one of the fields on a form. Alternatively, it could have been too costly or time-consuming to collect that particular feature for that particular instance.
BigML can handle missing data as input to obtain predictions and also in your training data to build models. Generally, missing values are treated as a unique value in itself, not mapped to any other value (either 0, the mean, or whatever). For instance, when using a decision tree model to compute predictions, you can choose what kind of strategy suits you best if a missing value is found in your input data. Two strategies are available: last prediction or proportional.
In the Last prediction strategy (chosen by default), your prediction is computed by descending the branches of the tree according to the values of your input data, until there's a question regarding your missing data. At this point the process stops and the prediction of the last node is returned. If you choose Proportional strategy instead (see image below), once you find a question that involves your missing data, the prediction algorithm does not stop but considers both branches as possible, and finally computes a prediction based on the majority (for classification) or the average (for regression) of all possible predictions found.
You can also enable this same option when using the BigML API. Please read this subsection of the documentation for developers.