Interference alignment (IA) is a widely recognized approach for mitigating inter-cell interference in multi-user multiple-input multiple-output (MIMO) networks. Despite its effectiveness, practical deployment remains constrained by two major challenges, i.e., the need for global channel state information (CSI) at each transmitter and the complexity of deriving closed-form solutions for intricate MIMO systems. This work aims to maximize network throughput by effectively mitigating interference using an IA-inspired learning algorithm that addresses its aforementioned challenges. First, we propose a predictive, transformer-based IA framework that estimates CSI to reduce signaling overhead in small-scale MIMO systems. Next, we formulate the IA problem as a multi-objective optimization problem based on subspace coordination and develop two reinforcement learning-based algorithms to enhance the scalability of IA in large-scale MIMO systems. Simulation results demonstrate that the proposed methods significantly outperform conventional baselines with up to 30% average user throughput gains over the best performing baseline.