Recently, significant progress has been made in sequential recommendation with deep learning. Inspired by neural language models, the standard approach to training sequential models is to use the user’s past behavior sequence as input and the user’s subsequent behavior as a supervisory signal. However, such a training procedure is fragile if the next immediate behavior in the training data is not correlated with the sequence of behaviors that occurred before this new behavior. This sequence mutation issue reflects the possible gap between past interacted items and target items, which inevitably causes confusion among the semantic information of the item embeddings. What’s even worse, optimizing sequential recommendation models on such a learning paradigm will result in many more wrong predictions, which significantly degrade the performance of the recommendation. In this paper, we formulate the existing recommendation models as a causal graph that reflects the cause-effect factors in sequential recommendation, and address the sequence mutation issue by building a new causal graph that can distinguish the effects of the user’s current intent and past interacted items on the prediction. To this end, we remove part of the direct effect from user features to prediction through counterfactual inference. By estimating the click likelihood of items in the counterfactual world, we are able to reduce part of the direct effect of user features and eliminate the noise issue. Experiments on real-world datasets demonstrate that our method improves the existing sequential recommendation models significantly.