Entropy Energy-Based Model: A neighborhood entropy energy-based model for high-dimensional data distribution estimation in intelligent communication networks

Abstract

Energy-based models have shown strong potential in unsupervised learning due to their flexibility in representing complex and uncertain distributions. However, for high-dimensional and locally non-uniform distributions that frequently occur in intelligent communication networks, traditional energy-based models face challenges such as the curse of dimensionality and data sparsity. In this paper, we propose a Neighborhood Entropy Energy-Based Model (NEEM), which estimates the non-uniform probability distribution of the k-Nearest Neighbor region around each sample point using a Local Gaussian. The energy function maps sample points to energy values for entropy estimation, enabling a more accurate representation of the data. NEEM is further applied to generative adversarial models to define the joint distributions of sample points, and its energy values are used for out-of-distribution detection. Experiments on benchmark datasets demonstrate the effectiveness of NEEM across seven evaluation metrics. The model achieves superior performance in diversity evaluation with a Fréchet Inception Distance of 28.88 and an Inception Score of 7.30. Results on out-of-distribution detection further show that NEEM outperforms five baseline models, highlighting its capability in high-dimensional probability distribution estimation and its potential for data analysis and anomaly detection in intelligent communication networks.