Abstract
Efficient formation of user and/or item groups is crucial for maximising individual user satisfaction across various recommendation scenarios. Most of the existing models employ explicit clustering techniques to construct such groups. Although these methods provide a straightforward way of group formation, they fail to adapt to the dynamic nature of real-world groups. As a result, these systems often produce suboptimal recommendations, particularly in contexts where group composition and item relevance are constantly evolving. To address this, we propose a novel Deep Dynamic Group Learning model (DDGLM) that dynamically learns latent group structures for both users and items within a unified neural architecture. Unlike conventional matrix factorization approaches, the model introduces probabilistic soft group assignments using a temperature-scaled softmax over user- and item-specific logits. These group probabilities are projected through linear transformation layers to obtain group-aware representations of users and items. The model supports a wide range of recommendation settings by learning both personalized and group-level representations without relying on static group definitions. It handles both scalar and ordinal prediction tasks through mean squared error and smooth hinge loss, respectively. Experimental results show that the proposed approach effectively captures latent group dynamics and consistently outperforms traditional group-aware baselines across multiple recommendation scenarios.