Markov Logic Networks With Its Application In Hypertext Classification And Link Prediction

In the real world,many problems are characterized by the presence of both uncertainty and complex.Probabilistic graphic models help us effectively handle uncertainty,while first-order logic help us compactly represent a wide variety of knowledge and relationships to reduce complexity. Combining the two in a single representation has been a research focus of Artificial Intelligence (AI) research field.Statistical Ralational Learning (SRL) approach combines probabilistic graphical model and the first-order logic together.SRL is the combination of relational (logic) representation,probabilistic inference (uncertainty processing) and machine learning (data mining), which aims to obtain the likelihood model from multi-relational data.Markov Logic Networks (MLNs) is a new Statistical Relational Learning model in which Markov networks and the first-order logic combine together. In traditional multi-relational data mining, a standard approach is to treat each entity identically,independent and identically distributed(IID), We use Markov Logic Networks in Hypertext Classification in order to solve this problem.The traditional approach ignore the potential structural information of object itself and the correlations between objects. Markov logic networks as a hot research field of artificial intelligence has been successfully used in the semantic role labeling, transfer learning, information extraction, molecular biology and other fields.This thesis focuses on applying the Markov logic network in Hypertext Classification and Link Prediction. The main research summarized as follows:Firstly, the thesis studies the related theory of Markov Logic Networks. This thesis firstly introdues the theoretical basis of Markov logic networks, which is first-order logic probability graph model and Markov network.And then ,we study the concepts of Markov logic networks .At last,we disscuss the weight learning algorithm and inference algorithm of Markov logic networks.Secondly, the thesis studies hypertext classification based on Markov logic networks.In traditional Hypertext Classification, a standard approach is to classify each entity independently, ignoring the correlations between them. We use Markov Logic Networks in Hypertext Classification in order to solve this problem. The experiments use discriminative learning method for Markov Logic Networks weights, and Gibbs sampling, simulated tempering, MC-SAT, and belief propagation algorithm for inference, which proves that this model has better performance than KNN in Hypertext Classification and the correlations between the entities are benefit for the performance as well.Thirdly, the thesis studies Link Predication based on Markov logic networks. Link prediction is an important and complex task, which aims to predict the relationships between entities. The traditional identically,independent and identically distributed model brings a lot of noise, resulting in bad performance. In order to solve this problem, we use Markov logic networks in Link Prediction. We focus on applying Markov Logic Networks to build a relational model, so as to predict the existence and the type of links between entities. Experiments results based on applying Markov Logic Networks model to two datasets prove Markov Logic Networks has better performance than traditional Link Prediction models, which providing evidences for Markov Logic Networks solving practical problems as well.To sum up: Markov logic networks is a powerful SRL mothed which combines probability and first-order logic. Markov logic networks not only handle uncertainty and complexity effectively, but also can be used as a template of constructing Markov network, which means it has a very wide range of applications.