Training Recurrent Neural Network on Distributed Representation Space for Session-based Recommendation

Recurrent neural networks (RNNs) have shown outstanding performance in the domain of session-based recommender system in the recent years. Previously these models are jointly trained with an item embedding layer to model the relations between items to make better recommendations. Nevertheless, bad representations of the items are acquired since the exploding gradient problem may arise during the training. In this work, we propose a novel method, where the representations are given by a shallow skip-gram model. Then we utilize the given representations to make recommendations with RNNs. Experimental results on the same dataset indicate not only relative improvements of about 6.6% and about 10.5% over previously reported results on Reca11@20 and Mean Reciprocal Rank@20 metrics respectively but also greatly reduce the training time.