ALGIN: Adaptive Local and Global Interests Network for Click-Through Rate Prediction

User behavior sequence modeling represents users' interests and preferences, which has a crucial impact on click-through rate (CTR) prediction models and recommendation systems. In practical scenarios, multiple types of interactive behaviors usually constitute users' complex interests and preferences. However, existing CTR prediction models based on multi-type interactive behavior modeling have two limitations: (1) they cannot handle noise in users' global interests, resulting in the inability to model accurate global interest preferences; (2) they struggle to depict users' local interest patterns and cannot model the complex patterns embedded in multi-behavior sequences. To overcome these limitations, an Adaptive Local and Global Interests Network For Click-Through Rate Prediction (ALGIN) is proposed, which effectively models global and local interests within multi-behavior interaction sequences through a unified framework. Firstly, ALGIN employs a Fourier Global Interest Modeling module to eliminate noise in the behavioral sequence; secondly, ALGIN designs a Multi-Scale Local Interest Extraction module to capture behavioral patterns within multi-type interactive sequences; finally, an Item-driven Interest Selector is used to integrate users' local interests with their global interests. Experiments on multiple public datasets demonstrate that the ALGIN model outperforms the current state-of-the-art (SOTA), and ablation experiments confirm the effectiveness of each module within the model.