多视图有监督的LDA模型

Multi-view Supervised Latent Dirichlet Allocation

本文主要关注多视图数据的分类问题.考虑到集成分类方法可组合多个弱分类器构成一个强分类器,以及主题模型能学习复杂数据的语义表示,本文试图将集成学习思想引入主题模型中,以便同时学习多视图数据的分类规则和预测性语义特征.具体地,结合概率主题模型LDA模型和集成分类方法 Softmax混合模型,提出了一个多视图有监督的分类模型.基于变分EM方法,推导了该模型的参数估计算法.两个真实图像数据集上的实验结果表明了提出模型有较好的分类性能.

In the paper,w e mainly focus on classifition on multi-view data.Considering that ensemble methods can combine w eak classifiers to construct a strong classifier,and topic model can learn latent representations from complex data,w e try to introduce ensemble idea to topic model,such that predictive latent representation could be obtained and multi-view classifier could be learned.We propose multi-view supervised latent Dirichlet allocation（ multi-view s LDA） model by combining latent Dirichlet allocation model and the mixture of softmax model w hich is an ensemble classification model.M oreover,w e derive a parameter estimation algorithm of the proposed model based on variational expectation maximization（ EM）procedure.The experimental results on tw o real datasets show the effectiveness of the proposed model.