GMBLOF: A Machine Learning Algorithm of Novelty Detection Based on Local Outlier Factor

Abstract

Local Outlier Factor (LOF) algorithm is a typical machine learning algorithm and has good accuracy in novelty detection for detecting global and local outlier. However, the LOF algorithm requires to traverse the entire dataset when calculating the local outlier factor of new data object. Aiming at the problem of extra time overhead in traversing entire dataset, we studied the combination of LOF and Growing Neural Gas (GNG) network and proposed the Growing-Model-Based Local Outlier Factor (GMBLOF) algorithm. Firstly, the topology of normal dataset generated by GNG was regarded as distribution description of the entire normal dataset and realized transformation from large to small order of magnitude. Secondly, the correlation of each node extracted from topology formed growing model. Finally, GMBLOF jointly growing model was used to calculate local outlier factor of new data point. The result of comparison experiments between GMBLOF and state-of-the-art novelty detection algorithms shows that the proposed algorithm outperforms in both accuracy and execution time, and especially the accuracy on WBC dataset improves by 8.69 % to 93.42%, compared to the other algorithms' average accuracy of 84.73%.