Representation Learning of Tangled Key-Value Sequence Data for Early Classification