Anomaly detection via adaptive greedy model

摘要: Abstract Anomaly detection is one of the fundamental problems within diverse research areas and application domains. In comparison with most sparse representation based anomaly methods adopting a relaxation term sparsity via l1 norm, we propose an unsupervised method optimized adaptive greedy model on l0 norm constraint, which more accurate, robust in theory. Firstly for feature representation, concise space learned way stacked autoencoder network. We dictionary selection l2, 0 constraint to select optimal small subset training data construct condense dictionary, can improve accuracy reduce computational burden simultaneously. Finally, each testing sample reconstructed by anomalies are determined depending reconstruction scores accordingly. For optimization, forward-backward utilized optimize this nonconvex problem theoretical guarantee. Our proposed evaluated our real industrial dataset benchmark datasets, various experimental results demonstrate that comparable conventional supervised performs better than comparative methods.