Event2Vec: A Geometric Approach to Learning Composable Representations of Event Sequences