Augmenting Sequential Recommendation with Pseudo-Prior Items via Reversely Pre-training Transformer