In order to effectively cope with the complex challenges of user behavior and preference changes over time, this paper proposes an adaptive recommendation model ADERec based on a dual network structure, which realizes accurate capture of user behavior dynamics by deeply integrating the time-series information of user-item interactions. The core of the model lies in its adaptive mechanism, which can dynamically adjust the recommendation strategy based on the user’s historical behavioral patterns to match the user’s current and future preference changes. Specifically, ADERec first builds a fine-grained interval-aware module using timestamp and item category information to analyze the patterns of user interest evolution. Subsequently, the output of this module is used as a guide to adaptively select or adjust data augmentation methods to ensure that the recommended content both matches the user’s current interests and stimulates their potential needs. To further enhance the accuracy and personalization of the recommendations, this paper also introduces a dual network structure to model the user’s historical behavioral data and potential future interests, respectively. This design not only realizes the two-way flow of information, but also ensures a high degree of consistency between the recommendation logic and the user’s actual behavior, which effectively solves the problem of “interest drift” in the traditional recommendation system. Experimental results show that when tested on multiple real data sets, the ADERec model improves the recommendation accuracy by 3% to 5% compared with current advanced methods, which fully verifies its effectiveness and superiority in practical applications.