TA-MOE: Multi-View Mixture of Experts with Tail-Aware Gating for Long-Tailed Specific Emitter Identification

Abstract

Specific Emitter Identification (SEI) in real-world deployments suffers from severe long-tailed class distributions. Existing single network approaches allow head classes to dominate the shared feature space, while direct application of Mixtureof-Experts (MoE) models introduces further issues of representational homogeneity and gating bias toward majority classes. This paper proposes TA-MoE, a tail-aware Mixture-of-Experts network for long-tailed SEI. TA-MoE employs three heterogeneous experts effectively decoupling feature space competition. A class distribution-aware gating mechanism incorporates class frequency priors. Furthermore, a dual-layer regularization scheme jointly optimizes expert load balance and tail-class entropy. Experiments on real-world ADS-B and AIS datasets demonstrate that TA-MoE consistently outperforms prior methods in overall accuracy, macro-averaged F1 score, and tail-class recognition.