Robust Sequence Embedding for Recommendation

摘要

Sequential recommendation is a significant task that predicts the next items given user historical transaction sequences. It is often reduced to a multi-classification task with the historical sequence as the input, and the next item as the output class label. Sequence representation learning in the multi-classification task is of our main concern. The item frequency usually follows the long tail distribution in recommendation systems, which will lead to the imbalanced classification problem. This item imbalance poses a great challenge for sequence representation learning. In this paper, we propose a Robust Sequence Embedding method for the recommendation, RoSE for short. RoSE improves the recommendation performance from two perspectives. We propose a balanced k-plet sampling strategy to make each training batch balanced at the data level and propose the triplet constraint for each training sequence to make sure of balance and robust distribution in feature space at the algorithmic level. Comprehensive experiments are conducted on three benchmark datasets and RoSE shows promising results in the face of item imbalance.