中文摘要：

研究贝叶斯正则化的自组织映射神经网络（self-organizing map,SOM）聚类训练算法。根据正则化的思想,在SOM权值调整公式中引入反映网络权值复杂性的惩罚项,避免权值调整过程中出现过度拟合。利用贝叶斯推理获取权值调整公式中的最优超参数,使迭代训练过程中网络权值和输入样本的概率分布更趋于一致,达到提升SOM聚类结果的目的。在UCI数据集上的实验结果表明,与传统的SOM算法相比,该算法的聚类凝聚度平均提升了1.5倍,聚类的准确率亦有提高,聚类效果较好。

英文摘要：

The self-organizing map clustering algorithm using Bayesian regularization was studied. According to the idea of regu- larization, during the weight adjustment process, the penalty term that reflected the complexity of the network weights was added to the weight adjustment formula, thereby avoiding overfitting. Bayesian inference was used to obtain the optimal hyper parameters in the weight adjustment formula, so that the network weights distribution and input data probability distribution became more consistent during the iterative training, and the clustering effect was improved. Experimental results on UCI dataset show that compared with the traditional SOM algorithm, clustering cohesion level of the presented algorithm is 1.5 times higher on average, the accuracy of clustering is also improved, and the clustering effect is much better.