具有随机化输入的贝叶斯概率模型

摘要

针对概率线性回归模型存在采用单层结构的表示能力有限、训练过程中容易存在过拟合问题，提出具有随机化输入的贝叶斯概率模型。通过对模型增加随机化输入层，对输入数据进行随机化处理，将单层线性处理模型转化为两层非线性模型以增强模型表示能力；同时对模型参数加入高斯先验概率分布以提高模型的泛化能力。理论分析和实验结果表明，具有随机化输入的贝叶斯概率模型具有较优的分类性能和较好的泛化能力。%In the probabilistic linear regression model, for the problem that the expressing capability of using the single layer structure is limited and the over-fitting is easy to occur in the training process, this paper presents a Bayesian probabilistic model with randomized input. By increasing the randomized input layers for the model, processing the input data at random, we convert the single layer linear processing model to the double-layer non-linear one to improve the model’s expression capability, and meanwhile, put Gaussian prior probability distri-bution into the model’s parameters to advance its generalization ability. Theoretical analysis and experimental re-sults indicate that Bayesian probabilistic model with randomized input is of better classification performance and generalization capability.