Feature Set Embedding for Incomplete Data

We present a new learning strategy for classification problems in which train and/or test data suffer from missing features. In previous work, instances are represented as vectors from some feature sp