Detecting Abnormal Relation In Linked Data

In recent years,research on linked data has shifted from building new data set to improving the quality of existing data.As the most critical factor in the linked data,the quality of the relationship is particularly important.However,due to the low quality of source data,the poor construction method and the lack of detection and supervision,the quality of the association is facing considerable challenges.There are a number of screening studies for specific types of anomalous relationships,but they are lacking in breadth.There are also some general-purpose relational detection methods that are obviously not deep enough.In this context,this paper proposes and implements a common set of techniques and methods including loop elimination,level assignment,frequent pattern mining,D-S evidence theory,and recommendation of hypernyms based on collaborative filtering.This paper contains three complements,as abnormal hypernyms relationship detection,abnormal attribute relationship detection and deletion of missing relationship.Specific work as follows:1)The detection and elimination of the relationship between the upper and lower loop.We transformed linked data into graph structure and then decomposed into a number of subgraphs by Unicom.Then,the existing loops are found on these independent subgraphs based on heuristic search,and based on the hierarchical distribution of level information and semantic information Algorithm to eliminate unreliable edges in the loop.2)Detection and elimination of abnormal property relations.This study finds that there is a frequent pattern between the hyperbolicals of the correct relationship between the two ends of the entity,and the frequent pattern can be used to reversely detect the abnormal attribute relationship.In order to obtain the ephemera,a classification system and an attribute knowledge base are extracted from Wikipedia.The former is used to mine the hypernym patterns of the two ends of the relationship based on the Apriori algorithm,and the latter is used to obtain the external knowledge.Then based on D-S evidence theory combined with frequent pattern observation and external knowledge observation,the author puts forward that the credibility of the relationship can be assessed synthetically,and the attribute relationship with low credibility can be excluded.3)The relationship between missing and abnormal missing detection and complement.This article draws on the idea of social network referral staff,using the method based on collaborative filtering recommendation,find the association with missing complements.This method designs a similarity matrix based on Jaccard similarity and random walking similarity to find similar entities,and accelerates similar entity queries using two methods:entity filtering and entity filtering.Then,the hypernyms of the TOP-K similar entities are proposed for the current entity through the voting mechanism.Finally,we conduct a series of experiments on multiple datasets.The experimental results show that using this method,the clear loop anomalies work at84.7% and 91.9% of Probase and SEBase,respectively,under the condition of relatively high recall rate % Accuracy;the relationship of removing anomaly attribute achieved the accuracy rates of 86.8% and 91.3% respectively on Yago and Freebase;the accuracy of 85.1% and 90.0% of the missing relationship complementation work in Probase and SEBase respectively.