Detecting outlier pairs in complex network based on link structure and semantic relationship

•The differences between link and semantics are utilized to detect outlier pairs.•A k-step index algorithm is proposed to calculate the term weighting.•Frobenius norm and linear transformation are combined to rank the top-K differences.•Direct and indirect link relations are both considered in link structure model. In this paper, we propose an outlier pair detection method, called LSOutPair, which discovers the vast differences between link structure and semantic relationship. LSOutPair addresses three important challenges: (1) how can we measure the target object's link similarity among multi-typed objects and multi-typed relations? (2) how can we measure the semantic similarity using the short texts? (3) how can we find the objects’ maximum differences between link structure and semantic relationship? To tackle these challenges, LSOutPair applies three main techniques: (1) two matrices are used to store link similarity and semantic similarity, (2) a k-step index algorithm, which calculates the term weighting for each object, (3) applying the linear transformation of Frobenius norm to matrices can obtain the top-K outlier pairs. LSOutPair considers link and semantics in complex network simultaneously, which is a new attempt in data mining. Substantial experiments show that LSOutPair is very effective for outlier pair detection.