Sequence-Based Recommendation with Bidirectional LSTM Network

In modern recommendation systems, most methods often neglect the sequential relationship between items. So we propose a novel Sequence-based Recommendation model with Bidirectional Long Short-Term Memory neural network (BiLSTM4Rec) which can capture the sequential feature of items to predict what a user will choose next. By collecting consumed items of a user in a sequence with time ascending order, fitting the model with the last item as the label, the rest items as the features, we regard this recommendation assignment as a super multiple classification task. Once trained well, the output layer of our model will export the probabilities of the next items with given sequence. In the experiments, we compare our approach with several commonly used recommendation methods on a real-world dataset. Experimental results indicate that our sequence-based recommender can perform well for short-term interest prediction on a sparse, large dataset.