Emotion-modulated non-stationary latent transition for multimodal sequential recommendation

User interest modeling for multimodal content distribution increasingly operates in non-stationary environments, where user preferences and behavioral patterns evolve over time and are strongly influenced by affective context. Existing sequential and emotion-aware recommendation methods typically treat emotion as an auxiliary feature or rely on stationary transition operators, limiting their robustness under preference drift. In this paper, we propose Emotion-Modulated Dynamic Interest Modeling (EM-DIM), which represents user interest as a latent state process whose transition dynamics are adaptively modulated by time-varying emotional states. By isolating emotional influence within the latent transition mechanism, EM-DIM enables regime-dependent interest evolution while remaining compatible with multimodal semantic encoding and incomplete affect signals. Extensive experiments on news, short-video, and e-commerce datasets demonstrate that EM-DIM consistently outperforms strong static, sequential, and multimodal baselines. On MIND, EM-DIM improves AUC from 0.704 to 0.714 and F1 from 0.401 to 0.417 over the strongest stationary sequential baseline. On KuaiRand, it reduces RMSE from 0.872 to 0.846 (3.0% relative reduction), and on Amazon from 1.041 to 1.018 (2.2% relative reduction). Further analyses show improved robustness under temporal drift and statistically significant user-level gains. These results indicate that emotion-modulated non-stationary dynamics provide a principled and effective foundation for reliable multimodal recommendation.