A Design Of Graph-based Anomaly Detection Framework For E-commerce Transaction Data

In recent years,E-commerce transactions have a fast increase as the development of the Internet.However,the development brings not only profits and convenience,but also abnormal behavior like account theft and fraud that threatens the benefits and safety of companies and users.Therefore,a proper user modeling method is highly desired to help discover existing abnormal operations and prevent potential losses,while the records of user event sequences can help a lot.A user event sequence consists of many operation events ordered by time which contain attributes pertaining to address,device and the operation itself.The traditional anomaly detection strategies which use handcrafted features based on prior observations are expensive and limited by the capability of experts.Besides,they also suffer from the limited and unbalanced labels.To address the challenges,we propose a graph based framework to model users' behavior and detect abnormal users given the event records of them.We first generate an attribute graph from the user event sequences to capture the relations between attributes inter-and intra-events.Then we compute attribute embedding based on the attribute graph and user event sequences as the initial input of user graphs.After that,we generate user graphs which represent user behavior in a way similar to the attribute graph,and adjust weights of nodes and edges on user graphs considering both user event sequences and the attribute graph.Finally,we apply Graph Neural Network to do graph classification for the adjusted user graphs to detect abnormal users.We conduct experiments on the real-world dataset with limited abnormal labels and the results show that compared with the best performance of other methods,the proposed method achieves an increase of 1.3% and1.8% of F1-score in two datasets and gets state-of-the-art performance in the anomaly detection task.We conduct further exploration and analysis on the generated graph and embedding,and the results demonstrate the effectiveness and rationality of our proposed approach.