关键词: 不平衡数据, 采空区自然发火, 支持向量机(SVM), 预测, h维空间的过采样算法(HSMOTE)

Abstract: In order to improve the prediction accuracy of spontaneous combustion of a few samples under the unbalanced data, a prediction model based on K-means-Relief-HSMOTE-SVM was built. First, the K-means method was used to optimize the traditional Relief method for index selection, to make up its deficiency-combustion indexes' unreasonablly high weights caused by the feature extraction under unbalanced data. Then, in view of problems such as overfitting encountered in dealing with unbalanced data with the SMOTE method an h dimensional spherical space thought was introduced, and the clustering algorithm was used to determine the center and establish the spherical space, and an improved HSMOTE algorithm was developed for balancing the spontaneous combustion data. Next, the SVM was used to predict the spontaneous combustion data. Actual samples from Xuandong No.2 coal mine were used to conduct 50 experiments, and a result comparison was made between the model built by the authors, the traditional SVM and other models. The results indicate that K-means-Relief-HSMOTE-SVM can effectively extract feature factors and overcome SMOTE defect,that compared with other models, K-means-Relief-HSMOTE-SVM can more effectively improve the traditional SVM in unbalanced prediction accuracy and geometric mean correct rate for combustion samples of the minority class natural data.

Key words: unbalanced data, spontaneous combustion in caving zone, support vector machine(SVM), prediction, h dimensional synthetic minority over sampling technique(HSMOTE)