Enhanced Adversarial Sequential Recommendation Fused with Siamese Networks for Edge Environments

With the rapid development of the Internet of Things (IoT), massive user behavior data are continuously generated at the network edge, making intelligent service optimization under constrained computational resources and dynamic environments a critical challenge. As a representative AI-driven service, sequential recommendation predicts users’ future preferences by modeling historical behavior sequences. Existing Generative Adversarial Network (GAN)-based sequential recommendation methods suffer from the discriminator’s hard-boundary classification, leading to feature misjudgment. Moreover, Transformer-based dot-product attention is sensitive to uncertain and noisy data, while window truncation limits long-range dependency modeling. Therefore, we propose an <strong>E</strong>nhanced <strong>A</strong>dversarial Sequential <strong>R</strong>ecommendation fused with <strong>S</strong>iamese <strong>N</strong>etworks for Edge Environments (EASNRec). With a lightweight architecture and robust feature modeling, the method is suitable for edge deployment under resource constraints and noisy conditions. First, a Siamese network is introduced into the GAN framework to project real and generated sequences into a unified feature space, where a similarity loss is constructed to jointly optimize the adversarial learning process of the generator and discriminator. Second, EASNRec integrates dynamic random augmentation with contrastive learning to explore the latent structure of user behavior sequences from multiple views, thereby enhancing model robustness. Furthermore, we design an <strong>U</strong>ncertainty <strong>A</strong>dversarial <strong>M</strong>emory <strong>M</strong>echanism (UAMM) that enables the model to suppress high-uncertainty and noisy data while capturing long-termdependencies in user behavior sequences. Experimental results show that EASNRec significantly outperforms state-of-the-art sequential recommendation models on three public datasets.