Edge Information Assisted Decoder for Business Process Anomaly Detection

Anomaly detection as a subject focuses on the identification of data point which significantly deviate from what is the norm or the standard of the dataset. This gives anomaly detection a wide range of applications where the detection of irregularities is often times of crucial importance such as Business Process Management (BPM). In this study we present a novel type of decoder referred to as “Edge Information Assisted Decoder” (EIAD), working on graph data to incorporate edge indexes and attributes into the decoding to achieve improved anomaly detection. We tested a total of 8 encoder-decoder combinations to comparatively evaluate them and prove the effectiveness of the proposed method. The proposed method and the best encoder-decoder combination, the graph attention network (GAT) encoder and the edge-conditioned convolution (ECC) decoder yielded an increase of 0.31 in F1-score from 0.32 to 0.63 when compared to the baseline multi-layer perceptron (MLP) decoder model, both with the optimal optimizer. The empirical results show that the proposed approach has a potential to improve graph based anomaly detection.