基于深度高斯过程的多元类别数据分布估计

Multivariate Categorical Data Distribution Estimation Based on Deep Gaussian Process

多元类别数据的可能取值会随向量长度的增长呈指数级增长,从而造成数据稀疏性问题。通过将观察数据嵌入到连续空间中训练识别数据之间的相似性,构建多元类别数据的线性高斯隐变量模型和类别隐高斯过程(CLGP)。在CLGP模型基础上,建立小样本多元类别数据分布估计的多元类别深度隐高斯过程模型,并结合蒙特卡洛采样的变分推断方法对模型进行参数优化。实验结果表明,与CLGP模型相比,该模型分布估计精确度有所提升。

The possible value of multivariate categorical data increases exponentially with the length of the vector,resulting in data sparsity.The similarity between the identified data is trained by embedding the observation data into the continuous space,and the linear Gaussian hidden variable model and the Categorical Latent Gaussian Process (CLGP) of the multi-category data are constructed.Based on the CLGP model,a multi-class deep hidden Gaussian process model for small sample multi-class data distribution estimation is proposed,and the parameters are optimized by Monte Carlo sampling.Experimental results show that compared with the CLGP model,this model distribution estimation accuracy has improved.