Research On Link Prediction Based On Time-Varying Networks

Link prediction is a branch of the link mining. According to attributes of theobjects and other observed links, link prediction is the problem of predicting theexistence of a link between two objects. The algorithms of link prediction can beused to extract the missing information, to predict the incidents occurring in thefuture and to evaluate the mechanism of the evolution of networks. The researchof link prediction plays an important role in the current popular applications ofsocial networks. For example, link prediction plays an irreplaceable role inpredicting the missing messages of social networks. Researchers in artificialintelligence and data mining have argued that a large organization, such as acompany, can benefit from the interactions within the informal social networkamong its members. Effective methods for link prediction could be used toanalyze social networks to get some credible conclusions.A key challenge for traditional data mining is tackling the problem of miningrichly structured, heterogeneous, non-law networks. Naively applying traditionaldata mining approaches based on the IID assumption can lead to inappropriateconclusions about the data. So, when it comes to these networks, we must takeaccount of potential correlations due to links and the relationship between objectschanging over time, for getting the appropriate mining results. In fact, objectlinkage is also knowledge that should be exploited when other knowledge ismined from networks. So, Time-Varying Network Model is proposed to quantifythe relationship between objects in this paper. The traditional algorithms of linkprediction are improved to be suitable for Time-Varying Network Model. Andone novel algorithm of link prediction based on Markov Logic Network isproposed in this paper. According to the experiment results on Enron dataset, theimproved algorithms of link prediction and the novel algorithm of link predictionare both better than the traditional algorithms of link prediction.In this paper, my main research topics include the following parts.①Time-Varying Network Model for social networks is proposed in thispaper. The traditional static graph of time varying network is just a simplerepresentation of events occurred between objects, rather than quantifies theprocess of relations between objects changing over time precisely. But Time-Varying Network Model not only includes the expression of the static graphof time varying network, but also introduces a time series that have a profoundimpact on the relationship between objects changing over time.②The traditional algorithms of link prediction are improved to be suitablefor Time-Varying Network Model. The improved algorithms of link predictionfor the time-varying network link prediction accuracy have significantlyimproved.③According to the attributes of Markov Logic Network, a novel linkprediction algorithm is proposed in this paper. We propose the novel algorithm oflink prediction that combines Markov Logic Network with the traditionalalgorithms of link prediction, for the reason that the results of traditionalalgorithms of link prediction in the dataset of different nature are very different,and even the results of different algorithms in the same model are diametricallyopposed, however, Markov Logic Network can be compatible with differentalgorithms of link prediction and even the exclusive algorithms. The novelalgorithm of link prediction for the time-varying network based on Markov LogicNetwork and the traditional algorithms of link prediction is much better than thetraditional algorithms of link prediction.