Neural Hidden Markov Model

Abstract

Hidden Markov models are tractable to capture long-term dependencies but intractable to compute the transition probabilities of higher-order process. We propose a neural hidden Markov models to compute the transition probabilities of higher-order hidden Markov model by a neural network and reduce the cost of computation. It is applied for time-aware recommender systems to show the benefits from the hybrid of combining neural network and hidden Markov model. We implement the recommender system and experiment on real datasets to demonstrate better performances over the existing recommender systems.