基于互信息的输入变量选择算法综述

空间环境是空间态势感知的重要组成部分，在空间科学领域，受限于对日地系统物理规律的全面认知，现有的基于物理机制的空间环境预测模型目前还难以实用化，通常采用的多是基于数据的经验类模型。基于互信息的输入变量选择算法为空间环境要素预测模型的输入确定了思路。由于能充分考虑不同输入、输入与输出之间的潜在关系，基于互信息的输入变量选择算法近年来在回归及分类问题上得到了广泛的应用和发展。本文以输入变量选择算法的3个关键环节，即评价标准、搜索策略和停止准则为线索，从不同角度对基于互信息的过滤式变量选择算法进行了系统的分析与梳理，重点对不同变量评价标准依赖的假设条件进行了数学上的推导和说明。最后总结了其发展规律，可为后续研究尤其是建立空间环境预测模型提供借鉴。

Overview of the Mutual Information-Based Input Variable Selection Method

The space environment is an important part of space situational awareness. In the field of space science, due to the incomplete understanding of the physics laws of the solar-terrestrial system, some existing space environmental prediction models based on physical mechanisms are still difficult to practicalize. It is more common to use the data-based empirical models. The input variable selection(IVS) algorithm based on mutual information(MI) can help to determine the impact inputs of the spatial environmental prediction models. Since taking into account of the potential relationship between different inputs, and between inputs and outputs, the MI-IVS algorithms have been widely applied to the regression and classification problems in recent years. From different perspectives, this paper systematically analyzes the current widely-used filter-based MI-IVS algorithms with the three key points of the IVS algorithms, namely the evaluation criteria, the search strategy and the stopping criterion as a clue. It focuses on the mathematical derivation of the assumptions of these criterias. Finally, the trend of the MI-IVS algorithms is summarized, which can provide reference for subsequent research, especially for establishing space environmental prediction models.