Identifying inconsistencies in object similarities from multiple information sources

A computer program product for detecting the horizontal anomaly, the computer program product comprising a non-transitory computer readable storage medium having computer readable program code embodied therewith for performing a horizontal anomaly detection method comprising:

• receiving a plurality of descriptions describing a plurality of objects, each object of the plurality of objects being described by a plurality of different information sources, wherein each individual information source of the plurality of information sources captures a plurality of similarity relationships between the plurality of objects;
• generating a similarity matrix from the plurality of different information sources, the similarity matrix comprising a feature vector of entries for each of the plurality of objects;
• and identifying at least one horizontal anomaly within the plurality of objects from the similarity matrix, wherein the horizontal anomalies each comprise a clustering of at least two objects of the plurality of objects into a common cluster based on a first information source of the plurality of different information sources and simultaneously clustering the at least two objects of the plurality of objects of the plurality of objects into different clusters based on a second information source of the plurality of different information sources, and wherein combining the information sources comprises;
• placing each individual similarity matrix along a block diagonal of the similarity matrix;
• and filling off-diagonal entries of the similarity matrix using weighted identity matrices, wherein a weight of the weighted identity matrices is a constraint on relationships across the plurality of information sources.

A horizontal anomaly detection method comprising:

• determining a plurality of N×N similarity matrices for N objects, wherein entries of each similarity matrix define similarities between different pairs of the N objects from a plurality of different information sources;
• performing a soft clustering on the similarity matrices together with a constraint that an object should be similarly assigned to a same cluster in each of the similarity matrices;
• and quantifying a difference between clustering solutions of each similarity matrices to identify scores that signify a clustering of at least two objects of the N objects into a common cluster based on a first information source of the plurality of different information sources and simultaneously clustering the at least two objects of the N objects into different clusters based on a second information source of the plurality of different information sources, wherein combining the information sources comprises;
• placing each individual similarity matrix along a block diagonal of the similarity matrix;
• and filling off-diagonal entries of the similarity matrix using weighted identity matrices, wherein a weight of the weighted identity matrices is a constraint on relationships across the plurality of information sources.