An Incremental Learning Method For Hierarchical Latent Class Models

Latent class models can be an important method for cluster analysis, the latent variables are used to represent the correlations between the manifest variables, thus can effectively handle the inherent information of the observed data set, and the method is practical in many applications nowadays. There are a lot of ways to learn the models involving latent variables, one of the most popular methods is the Hierarchical Latent Class models that expand the traditional Latent Class models, it can perform multidimensional clustering analysis on the observed data, and can genuinely reflect the intrinsic relations among the observed data.However, when the Hierarchical Latent Class models are constructed based on large amount of data samples, there are two problems remaining unresolved. Firstly, we need to introduce several latent variables to represent the correlations between the manifest variables, so the EM algorithm has to be used to estimate the network parameters that include missing data, which results in the computationally complexity and time consuming. Secondly, when the model structure is updated, too many candidate models are constructed for next refinement process, then we need too much storage capacity to store these candidate models. Therefore, it's not realistic for us to use the traditional batch method, which means the new data and old data are combined as the whole data set, to construct models every time the new data are observed.For solving these problems, a new incremental learning method in the context of Hierarchical Latent Class models is proposed in this thesis, The focus of the work is twofold:●The first is to identify the influenced latent structures for the existent Hierarchical Latent Class model.According to the Maximum Likelihood Estimation (MLE), we define a data coincidence degree for each latent node to measure the influencement taken by new data, and it is useful to compress the learnt knowledge.●The second is to adopt the Markov Blanket as the recovering radius to partition the network, then we can refine sub networks in a small range. These can not only utilize the already learnt knowledge, but also solve the problem that the memory is limited by the large amount of data.The experimental results show that the method is feasible, and has advantages than the traditional algorithm for learning Hierarchical Latent Class models.