<p>Multimodal recommendation (MMRec) utilizes multimodal content of positive items (that users have interacted with) to determine user preferences, while ignoring the rich information in negative items (that users have not interacted with). Recently, some counterfactual-learning-based recommendation have leveraged causal differences between positive items and negative items to improve recommendation performance. However, these methods focus heavily on improving item representations, while neglecting the modeling of user preferences, making it difficult to capture genuine user preferences. Therefore, we propose a model called Counterfactual Enhanced Recommendation(CFER), aiming to improve the quality of modeling user preferences while utilizing negative item information. First, we capture fine-grained differences between positive and negative item using counterfactual inference combined with a disentanglement technique. We then use this difference to construct a weight distribution for the user interests and re-weight the target user representation, thereby highlighting the users’ true preferences and enhancing the user preference representation. We conduct experiments on the Beauty, Art, and Taobao datasets, and achieve average improvements of 0.83% and 0.99% over the state-of-the-art methods in terms of HR and NDCG, respectively.</p>