Relationship Minging In Heterogeneous Information Networks Based On Multi-label And Multi-instance Learning

Many complex systems in the real world can be formalized as networks where nodes represent objects and links represent interactions between objects.Most of these networks are heterogeneous,containing several types of objects and relationships.We take the online social network Twitter as an example,it contains nodes such as users,locations and tweets,and links such as posting tweets,forwarding tweets,following users,check-ins.As a key issue in link mining,link prediction predicts the formation of potential or invisible links based on current or historical networks.Previously,most of the methods of link prediction have been designed for homogenous information networks,where there is a single type of nodes and links in the network.Recently,link prediction research towards heterogeneous information networks has attracted increasing attention from the researchers,with some new progress made.In this context,we propose a relationship predictor MULRP based on multi-label learning and a link predictor MUIRP based multi-instance learning.For the first time,MULRP introduces multilabel learning into the field of relationship prediction.It uses meta-paths to define the relationships between nodes and treats them as labels and designs a new predictive relationship model in a multilabel learning framework.The experimental results show that this method can not only predict the new target relationship more accurately,but also reveal the correlation between different types of relations(i.e.tags)and provide suggestions on how to form new relationship.MUIRP treats the relation prediction as a multi-instance learning problem.We take feature vector of each relationship as an instance.A bag composed of multiple instances is used to represent a sample,if there is a target relationship between a node-pair,then the label of bag is positive,otherwise negative.Experimental results show that the performance of MUIRP is superior to the traditional binary supervised learning method but poorer than the MULRP method.