Research Of Multiview Sequence Data Modeling Based On Conditional Random Fields

Multiview learning has gained much attention due to the increasing amount of multiview data,which are collected from different sources or can be characterized by different types of features.Therefore,this paper mainly proposed multiview conditional random field models and algorithms.This paper proposes two multiview representation models(acquiring multiview representations)and two multiview conditional random field models(modeling multiview sequential dynamics)to model multiview sequential data.On the one hand,we want the processed multiview representation has good properties.Considering some unique characteristics like heterogeneity,complementarity,and redundancy in multiview data,we try to decouple and normalize the data to obtain a unified multiview representation,which can facilitate modeling after.Thus,based on the feature engineering and representation learning,we propose two multiview decomposition representation models: multiview feature recombination conditional random field and neural multiview decomposition network.On the other hand,based on the unified multiview representation,we introduce a hidden variable to balance the usage of the model in different views.We believe that the information contained in the multiview sequential data is dynamically changing over time.Specifically,the model focuses on different views at a different moment with varying degrees.In order to calculate the expectation with the latent variables more efficiently for different views,we proposed two different mechanisms to approxoamte the latent variable based on the conditional random field: latent weight variable conditional random field and soft weight condition random field.In the model learning of latent weight variable conditional random field,we approximate posterior with tighter approximation bound based on variational inference.Furthermore,we introduce a method of controlling variables to reduce the variance of the gradient to make the model calculation more feasible.In the soft weight CRF,we used a deterministic weight to approximate the latent weight variable,which make the computation feasible.In summary,the multiview conditional random field proposed in this paper mainly includes two components: weight variables condition random field and multiview decomposition representation model.The introduction of weight variables and the decomposition representation of multi-view data enable multiview conditional random fields to model multiview sequential data more reasonably.We experimented with two data sets in the real world to compare the proposed method with other related methods.The experimental results show that the proposed methods have performance that is more outstanding.At the same time,in order to verify the role of the two components in the multiview conditional random field,we also test and analyze with only the weight variable model and the multiview decomposition representation model.Experimental results and careful analysis show that both the proposed models contribute to the outstanding performance over other methods.