Automated design of tailored link prediction heuristics for applications in enterprise network security

Abstract

The link prediction problem, which involves determining the likelihood of a relationship between objects, has numerous applications in the areas of recommendation systems, social networking, anomaly detection, and others. A variety of link prediction techniques have been developed to improve predictive performance for different application domains. Selection of the appropriate link prediction heuristic is critical which demonstrates the need for tailored solutions. This work explores the use of hyper-heuristics to automate the selection and generation of customized link prediction algorithms. A genetic programming approach is used to evolve novel solutions from functionality present in existing techniques that exploit characteristics of a specific application to improve performance. Applications of this approach are tested using data from a real-world enterprise computer network to differentiate normal activity from randomly generated anomalous events. Results are presented that demonstrate the potential for the automated design of custom link prediction heuristics that improve upon the predictive capabilities of conventional methods.