GenSim : GAN based Recommendation systems for personalized matrix factorization

Accurately modeling user preferences is crucial for the success of modern recommendation systems (RS). Despite recent advances in generative models for RS, challenges such as limited data and the complexity of human behavior still persist. These issues make it difficult to generate accurate and authentic user profiles, which are essential for providing meaningful and personalized recommendations. In this paper, we introduce GenSim, a novel approach that combines Generative Adversarial Networks (GANs), Genetic Algorithms (GAs), and similarity techniques to overcome critical challenges in collaborative filtering (CF), such as data sparsity and intricate user-item interactions. By integrating these methods, GenSim offers a robust and scalable framework for enhancing RS performance. A key feature of GenSim is its focus on personalized matrices, which selectively consider only similar users or items, rather than the entire user-item matrix. This targeted approach refines the input data during the generation phase, resulting in recommendations that are not only more accurate leading to more accurate, personalized, and efficient recommendations. Our approach integrates GAN with an autoencoder-based discriminator and an optimized generator for matrix factorization, incorporating Pearson similarity data to enrich the generative process. GAs are employed in two phases: preprocessing to refine user similarity measures and fine-tuning the generator’s hyperparameters for optimal matrix factorization. Extensive experiments on benchmark datasets—MovieLens 1M, HetRec, and LastFM—demonstrate GenSim's superior performance across Precision, MAP, and NDCG, compared to state-of-the-art methods. Our approach improves precision by 40% at cutoff 50 for the MovieLens 1M dataset, and MAP by 38% at cutoff 5 for the LastFM dataset, compared to previous works using GANs for matrix factorization in recommendation systems.