CAJ | 학술논문

真实世界中不同类型的对象之间相互连接，形成异质信息网络．预测网络中对象之间的连接或交互是网络分析中的一个重要任务．不同于传统的同质性网络的链路预测，异质信息网络中，由于存在多种类型的节点和边，节点之间可以通过不同的关系进行连接．文中使用元路径，即通过一组关系连接了多种节点类型的路径，来描述异质信息网络中不同类型对象之间各种连接的不同语义，从而提出一种异质信息网络链路预测模型，通过组合对象之间在不同元路径上建立连接的概率来进行链路预测．在DBLP和Last．fm两个真实数据集上的实验结果表明：在7种关系的链路预测中，相比最好的基准方法，文中方法的AUC值平均提升了5．93％；另外，在链路预测中，通过元路径区分不同类型的节点和边之后，预测精度得到了明显提升；最后，为了平衡预测精度和模型的可扩展性，实验分析表明链路预测中仅考虑路径长度小于5的元路径就已经足够产生很好的预测结果．
진실세계중불동류형적대상지간상호련접，형성이질신식망락．예측망락중대상지간적련접혹교호시망락분석중적일개중요임무．불동우전통적동질성망락적련로예측，이질신식망락중，유우존재다충류형적절점화변，절점지간가이통과불동적관계진행련접．문중사용원로경，즉통과일조관계련접료다충절점류형적로경，래묘술이질신식망락중불동류형대상지간각충련접적불동어의，종이제출일충이질신식망락련로예측모형，통과조합대상지간재불동원로경상건립련접적개솔래진행련로예측．재DBLP화Last．fm량개진실수거집상적실험결과표명：재7충관계적련로예측중，상비최호적기준방법，문중방법적AUC치평균제승료5．93％；령외，재련로예측중，통과원로경구분불동류형적절점화변지후，예측정도득도료명현제승；최후，위료평형예측정도화모형적가확전성，실험분석표명련로예측중부고필로경장도소우5적원로경취이경족구산생흔호적예측결과．
Real-world,multiple-typed objects are often interconnected,forming heterogeneous information networks.The problem of predicting links or interactions between objects in a net-work is an important task of network analysis.Different from traditional link prediction tasks inhomogeneous networks,heterogeneous information networks have multiple distinct types ofinterrelated objects and relationships,and the objects in heterogeneous networks can be connectedvia different relationships.In this paper,we use meta-path,a kind of path that connects objecttypes via a sequence of relations,to describe distinct semantics of different types of nodes andrelationships,and propose a meta path-based link prediction model for heterogeneous information networks (MPLP).In this model we perform link prediction by combining the link probabilitiesbetween nodes on different meta-paths.The method is evaluated with two different genres ofdatasets:DBLP and Last.fm.Experimental results show that when predicting the formation ofseven types of relationships,the AUC value of our method is on average improved 5.93% compared with the most accurate method.Furthermore,distinguishing different types of nodesand relationships by using meta-paths can improve the prediction accuracy significantly.At last,in order to balance the prediction accuracy and the scalability of our model,an experimentalanalysis demonstrates that considering meta-paths whose lengths are less than 5 is enough toproduce good results in terms of link prediction.