Interpretability Analysis of End-to-End Learned CSI Feedback for Precoding-Oriented Massive MIMO

Recent end-to-end learning frameworks jointly optimize pilot design, channel state information (CSI) feedback, and downlink precoding, achieving strong performance under limited feedback overhead in massive multiple-input multiple-output (MIMO) networks. However, the internal representations and decision-making strategies learned by these architectures remain largely unexplored. This paper presents an interpretability analysis of a precoding-oriented end-to-end CSI feedback framework using explainable artificial intelligence (xAI) techniques. Building on a validated learned architecture, we investigate the structure of the learned pilots, latent feedback representations, and precoding vectors through low-dimensional visualization, density-based clustering, and cross-block representation analysis. The results reveal that the learned system undergoes a transition as the feedback constraint is relaxed. Under stringent feedback budgets, the encoder develops clustered latent representations and quantized precoding behavior. As additional feedback resources become available, the learned representations evolve into smooth geometry-preserving manifolds that retain physically meaningful channel characteristics such as dominant angular information, while enabling more adaptive precoding decisions. Furthermore, the learned pilot signals naturally form complementary sensing patterns aligned with the underlying channel statistics. The analysis also reveals a shift from gain-driven beamforming strategies under severe compression to increasingly interference-aware precoding as the feedback budget increases. These findings provide new insight into how end-to-end learned CSI feedback systems balance compression efficiency and communication performance and establish a general framework for interpreting learned communication architectures.