In e-commerce recommendation, user behavior sequences often contain strong, weak, and irrelevant interests concerning the target item. Existing sequential models tend to overfit weak interests, leading to biased predictions. To address this overfitting challenge, we propose the Adaptive Multi-Interest Counterfactual Network (AMIC-Net). It is designed to amplify strong interest contributions while mitigating weak interest overfitting and noise interference. Specifically, AMIC-Net employs: 1) An Implicit Multi-Interest Segmentation Module (IMSM) that implicitly segments user interests by sorting the item sequence based on continuous, attention-based relevance scores, before partitioning and evaluating the resulting segments. 2) An Adaptive Weighting Module (AWM) that uses the evaluation score of each segment as a dynamic weight, amplifying strong interest signals while attenuating weak and irrelevant ones to generate a core-intent-focused user representation. 3) A Counterfactual Fusion Module (CFM) applying counterfactual reasoning to integrate the base model’s predictions with AWM’s weighted interest representation. This fusion specifically enhances the direct predictive effect of strong interests, thereby reducing the overfitting risk associated with weak interests. Notably, our plug-and-play method easily integrates with existing models, and online A/B tests confirm that AMIC-Net improves recommendation quality in industrial applications. The implementation is available at https://github.com/SunChuike/AMIC-Net .