An adversarial framework with dual genetic optimization for similarity-aware matrix factorization in recommendation systems

Abstract

Modern recommendation systems (RS) continue to face critical challenges, including data sparsity and the difficulty of modeling complex user–item interactions, which hinder their ability to deliver accurate and personalized recommendations. To address these limitations, we propose GaSimGAN, a novel collaborative filtering framework that integrates Generative Adversarial Networks (GANs), similarity-aware modeling, and a dual genetic optimization strategy. The proposed framework leverages personalized similarity matrices to focus on the most relevant users and items, thereby refining the input space of the generative process and improving recommendation accuracy. GaSimGAN adopts a matrix factorization-based generator coupled with an autoencoder-based discriminator, enriched with Pearson similarity information to better capture relational patterns in user–item interactions.Genetic Algorithms are employed in a dual role: first, during preprocessing to optimize similarity-based neighbor selection, and second, as an offline one-shot hyperparameter optimization step conducted prior to and independently of the adversarial training pipeline. Extensive experiments conducted on benchmark datasets, including MovieLens 1M, HetRec2011, and LastFM, demonstrate that GaSimGAN consistently outperforms state-of-the-art GAN-based recommendation systems in terms of Precision, Mean Average Precision (MAP), and Normalized Discounted Cumulative Gain (NDCG), confirming both its effectiveness and scalability.