Research And Application Of Meta-graph Similarity Search Method In Heterogeneous Information Network

Similarity search in heterogeneous information network have been widely applied in data mining such as recommendation and anomaly detection.The main task is to find the most similar entities for the entities in the network.Meta-path-based similarity measurement have been attracted extensive attentions.Meta-graph proposed recently enhances the expressiveness than meta-path,but it is also a computing-expensive algorithm.Grounded on the above contradictions,this paper research on the correlation algorithm of similarity measurement based on meta-graph.The main research work of this paper could be concluded as follows:(1)A normalized similarity measure of meta-graph(NSMG)is proposed to balance the effect with computing cost.By taking advantage of the low-cost computing meta-path-based similarity measure,PathSim was extended to meta-graph to eliminate the bias of highly visible entities.NSMG adopts dot product in consideration of the information loss caused by the combination of meta-path.Compared with the current popular similarity measures such as PathSim,the experiments have shown that NSMG algorithm balances the rich semantics of meta-graph and low computing costs.(2)In the real scenario of review website Yelp and e-commerce website Amazon,a recommendation algorithm framework NSMG-Rec is proposed to recommend entities for users.The NSMG-Rec framework uses the NSMG algorithm to obtain the similar score matrix of "user-merchant/commodity",and uses the combination of matrix decomposition and factorization machine to predict the user ratings of merchant/commodity.The experiment not only compared the effects of NSMG-Rec and common recommendation algorithm,but also analyzed the effects of different meta-graphs.(3)In the real Weibo scenario,an algorithm framework NSMG-Anom is proposed to detect abnormal event.NSMG-Anom framework uses NSMG algorithm to add meta-graph similarity information,which effectively improves the performance of anomaly detection.Experiments show that NSMG-Anom algorithm can achieve better results than common anomaly detection algorithms in the detection of defined events,such as hack-time.