Label Semantic Expansion via Label Guided Neural Topic Modeling
为解决标签中心分析中标签表示稀疏问题,提出通过标签引导神经主题模型(LGNTM)来丰富标签语义,提高标签-主题对齐和分类性能。
为解决标签中心分析中标签表示稀疏问题,提出通过标签引导神经主题模型(LGNTM)来丰富标签语义,提高标签-主题对齐和分类性能。
本文针对现有医疗问答系统缺乏适应性、持久记忆和结构化决策的问题,提出了一种基于多代理的自适应记忆与反思系统,通过专门的记忆和反馈机制提高复杂案例处理能力。
This work addresses the challenge of reconstructing full-port channel state information (CSI) in fluid antenna systems, where hardware constraints permit observation of CSI from only a limited number of ports and prior spatial covariance statistics are unavailable. Leveraging the Clarke isotropic scattering model, the study reveals that the channel resides in a low-dimensional modal subspace dictated by the scattering environment. It establishes, for the first time, a theoretical framework for CSI recoverability without requiring prior statistical knowledge and derives an exact feasibility threshold. By integrating spatial modal modeling, error decomposition analysis, and low-rank recovery techniques, the proposed approach enables high-fidelity reconstruction of full-port CSI using only a few active ports, substantially reducing the number of required RF chains, pilot overhead, and training data.
研究解决了知识蒸馏中模仿准确性和生成鲁棒性的平衡问题,通过重新审视并改进RKL方法,提出自适应熵蒸馏(AED),动态调整学生模型的模仿强度。
This paper addresses portfolio optimization under investor preferences and margin trading mechanisms. Methodologically, it proposes a constraint-guided metaheuristic framework that: (1) explicitly incorporates pre-allocation constraints to encode investor preferences and reduce the feasible solution space; (2) endogenizes margin trading costs and leverage risk into a dynamic return model; and (3) replaces the conventional Sharpe ratio with the Rare Performance Ratio (RPR) to enhance robustness and practicality in risk-adjusted performance evaluation. The framework innovatively integrates pre-allocation constraints and leverage-aware strategies within standard metaheuristics—including genetic algorithms and particle swarm optimization—enabling synergistic constraint handling and search guidance. Empirical results across realistic market settings demonstrate statistically significant outperformance over traditional benchmarks, with an average 18.7% improvement in risk-adjusted returns. This validates both the methodological efficacy and operational deployability of the proposed approach in practical investment management.
为解决标签中心分析中标签表示稀疏问题,提出通过标签引导神经主题模型(LGNTM)来丰富标签语义,提高标签-主题对齐和分类性能。
本文针对现有医疗问答系统缺乏适应性、持久记忆和结构化决策的问题,提出了一种基于多代理的自适应记忆与反思系统,通过专门的记忆和反馈机制提高复杂案例处理能力。
This work addresses the challenge of reconstructing full-port channel state information (CSI) in fluid antenna systems, where hardware constraints permit observation of CSI from only a limited number of ports and prior spatial covariance statistics are unavailable. Leveraging the Clarke isotropic scattering model, the study reveals that the channel resides in a low-dimensional modal subspace dictated by the scattering environment. It establishes, for the first time, a theoretical framework for CSI recoverability without requiring prior statistical knowledge and derives an exact feasibility threshold. By integrating spatial modal modeling, error decomposition analysis, and low-rank recovery techniques, the proposed approach enables high-fidelity reconstruction of full-port CSI using only a few active ports, substantially reducing the number of required RF chains, pilot overhead, and training data.
研究解决了知识蒸馏中模仿准确性和生成鲁棒性的平衡问题,通过重新审视并改进RKL方法,提出自适应熵蒸馏(AED),动态调整学生模型的模仿强度。
This paper addresses portfolio optimization under investor preferences and margin trading mechanisms. Methodologically, it proposes a constraint-guided metaheuristic framework that: (1) explicitly incorporates pre-allocation constraints to encode investor preferences and reduce the feasible solution space; (2) endogenizes margin trading costs and leverage risk into a dynamic return model; and (3) replaces the conventional Sharpe ratio with the Rare Performance Ratio (RPR) to enhance robustness and practicality in risk-adjusted performance evaluation. The framework innovatively integrates pre-allocation constraints and leverage-aware strategies within standard metaheuristics—including genetic algorithms and particle swarm optimization—enabling synergistic constraint handling and search guidance. Empirical results across realistic market settings demonstrate statistically significant outperformance over traditional benchmarks, with an average 18.7% improvement in risk-adjusted returns. This validates both the methodological efficacy and operational deployability of the proposed approach in practical investment management.