This paper tackles the CTR prediction challenge within Trigger-Induced Recommendation (TIR) contexts. In TIR, users’ current interests are exposed to some extent through clicked trigger items, which then lead to the recommendation of preference items. Discovering decomposed trigger-relevant and trigger-irrelevant user interests and adaptively fusing them for final CTR prediction remains a challenging and under-explored area in existing TIR-based CTR methods. To address this challenge, we introduce the User Interest Decomposition Network (UID-Net), an innovative model consisting of three key modules: Intent Extraction Module (IEM), Interest Decomposition Module (IDM), and Interest Fusing Module (IFM). IEM initially decomposes each behavior’s embedding vector into trigger-relevant and trigger-irrelevant component embeddings, forming two component sequences used to extract corresponding interests. IDM employs self-supervised contrastive learning to further distinguish these interests, resulting in more discriminative representations. Finally, IFM predicts a fusion weight that adaptively combines trigger-relevant and trigger-irrelevant interests for accurate CTR prediction. Extensive offline-to-online experiments showcase the superiority of UID-Net to the state-of-the-art (SOTA) models.