The Research On Link Prediction Problem Based On Unifying Multisource And Heterogenous Data

In recent years,with the rapid growth of social network services and other network applications,social network data has shown an exponential growth trend,and these data begin to correlate and intersect with each other.As these data preserve the potential behavior patterns of users to some degree,how to extract the potential association of users from the massive social network data to improve the quality of social network services has become an urgent problem for many social networking sites and other related enterprises.At the meantime,it has also caused a research upsurge in the academic community.Use the way of completing social network,Link Prediction mine the potential business value of large data.Link Prediction is defined to predict the possibility of links between two nodes in a network that have not generated links yet through utilizing the information of known network nodes and network structure.Due to the data used by Link Prediction are multidimensional and comprehensive,there are often unexpected correlations between data.However,traditional Link Prediction does not dig deeply enough into these multi-dimensional data and ignore the relevance of these data.This thesis explores these multi-dimensional data layer by layer,creates new Link Prediction models,and achieves more accurate and efficient prediction results.In view of the limitations of traditional link prediction,this thesis constructs hybrid models of Link Prediction based on multi-source heterogeneous merged data through neural network.The main research contents are as follows:1.A hybrid Link Prediction model based on multi-source heterogeneous data is proposed in this thesis.The model uses two heterogeneous data which are from Location-based Social Network dataset,include user relationship topology map and user check-in records,to mine user behavior patterns.To some extent,this model improves the accuracy of traditional Link Prediction model.2.This thesis constructs a hybrid Link Prediction model based on Anchor Link algorithm.This model uses anchor link algorithm to exploit the potential association among multi-dimensional data of Location-based Social Network dataset.It captures the association between multi-source heterogeneous data sufficiently,and its prediction accuracy is better than the hybrid Link Prediction model based on multi-source heterogeneous data.3.A hybrid link prediction model based on Locality Sensitive Hashing technology is implemented in this thesis.The model uses Locality Sensitive Hashing technology to transform multi-source heterogeneous merged data which were extracted from Location-based Social Network dataset into Hamming code.On one hand,Locality Sensitive Hashing technology can preserve the similarity between the original associated nodes.On the other hand,using the Hamming code to train model can improve the computation speed and reduce data storage consumption.Compared with the two previous models,the performance and accuracy of this model are further improved.Compared with the traditional Link Prediction methods which using only one single data source,such as walk2 friends,node2vec and so on,three hybrid models of Link Prediction proposed in this thesis are tested on public datasets such as Gowalla and Foursquare,and utilize AUC,F1 value and other evaluation indicators to verify the feasibility and efficiency.The experimental results show that the hybrid models proposed in this thesis are more efficient and accurate than the traditional Link Prediction methods.