A More Examples of the Semantic Inconsistency Problem 1 (a) User A (b) User B

Each picture represents a news article clicked by the user. Dash borders indicate behaviors replaced by the augmentation method. The data augmentation proportion is set as 0.6. We also find that the behavior sequence augmented by masking well preserves the user's The pseudo-codes of the pre-training procedure with our AdaptSSR are shown in Algorithm 1. Randomly select two augmentation operators f and g from A. With independently sampled dropout masks. With independently sampled dropout masks.