Abstract
There exist two pervasive context conceptualizations (i.e. interactional and representational contexts) in sequential service recommendations. These approaches are typically limited in their ability to capture diverse transition patterns because they assume a single global context space for all sequences. In this work, we propose a context embedding model that augments an attention RNN encoder-decoder with a deep clustering component, to address the above limitation. The deep clustering component predicts a soft cluster assignment for the final encoder hidden state of each input sequence. The clustering assignment is subsequently used to condition both interactional and representational context and ultimately improve sequential prediction. We perform experiments to evaluate and validate our proposed model on four benchmark online services datasets. Our experiments show that our proposed model mostly outperforms state-of-the-art approaches on the sequential recommendation task.