A hybrid framework of LLM-based interaction generation and adaptive sampler for recommendation enhancement

Data augmentation (DA) is a critical technique for addressing data sparsity. Existing DA methods suffer from several limitations: extensive parameter tuning, dependency on specific types of data, and the tightly coupled DA network in the backbone recommendation model. Consequently, there remains a notable absence of universal data augmentation frameworks capable of adapting to diverse recommendation tasks and backbone models. To address these issues, we propose a hybrid framework to mitigate data sparsity via DA in a more flexible and universal paradigm. Notably, we disentangle the data augmentation process into two modules, data generation and data fusion. Firstly, we prompt LLMs to learn collaborative filtering information from raw user–item interaction data in a CSV format, generating new reliable interactions. Specifically, we devise clustering-based data grouping and multi-round shuffling strategies to enable LLMs to capture comprehensive collaborative filtering information through limited context. Secondly, we design an adaptive sampler to obtain personalized fused data for backbone recommendation models dynamically, making the fused data satisfy the different data demands of diverse downstream recommendation models. Experiments on various recommendation models across distinct datasets demonstrate that our framework can consistently improve vanilla recommendation models and facilitate existing DA methods. Compared with vanilla recommendation models, our framework achieves an average performance gain of 7.10%, 23.54%, and 24.55% on MovieLens-100k, Book-Crossing, and Toys and Games datasets, respectively. These results highlight the effectiveness of our proposed framework in mitigating data sparsity in recommender systems.