A subgraph-based knowledge reasoning method for collective fraud detection in E-commerce

Abstract

Fraud detection is essential for e-commerce platforms to maintain a fair business environment. Many existing works propose manually designed methods such as label propagation and dense block mining rules on built user-item graphs to detect fraud behaviours, but they are always heuristic and thus have limited performance. Other learning-based methods can either handle only the fraud detection problem well in the transductive scenario when there is only structural information or require rich content features to obtain a good inductive ability. Considering that content features are not always available in practice and there are usually many fraudulent behaviours that belong to newly emerging users and items, how to learn effective inductive rules with structures only is still underexplored. In this paper, we propose a subgraph-based method named SubGNN for collective fraud detection. In SubGNN, first, we extract the subgraphs around the given edges (user behaviours) to be tested. Then, we remove nodes’ global IDs so that SubGNN is entity-independent. Finally, by learning knowledge reasoning rules on extracted heterogeneous subgraphs using our proposed relational graph isomorphism network (R-GIN), a powerful graph neural network (GNN) model, SubGNN can achieve precise fraud detection. Experiments are conducted on publicly available Amazon and Yelp datasets and a newly collected Taobao dataset. The results clearly show the advantages and prospects of our method. When using SubGNN to detect fraudulent transactions on Taobao, the precision is higher than $0.99$ and more than $90\%$ of fraud samples are recalled.

Introduction

To better match users’ interests and items, various data-driven recommendation and ranking algorithms [1], [2] have been designed and deployed in existing e-commerce platforms. Although these algorithms can predict users’ interests better and bring more profits to the platform, the data-driven training mechanism makes them easily poisoned [3], [4]. For example, if a seller controls some user accounts to frequently click a popular item and a target item that belongs to the seller, the collaborative filtering-based recommendation algorithms may be misled that there are some important connections between these two items and begin to recommend the target item to the users who have interactions with the popular item, thus allocating much more impressions [5] to the target item than before.

Fortunately, only if malicious sellers control (or hire) some user accounts and perform a certain amount of fraud behaviours (log data) on real-world e-commerce platforms could the deployed recommendation or ranking algorithms in the system be misled. Considering that the number of accounts that malicious sellers can control is limited and that they always promote several target items together each time, these accounts’ behaviours are usually more collective than those of normal users [6]. This provides us with the opportunity to detect fraud behaviours.

There have been many effective solutions based on label propagation [7] and dense block assumptions [6], [8], [9], [10] for fraud detection in the literature. Similarly, these methods all first build a user-item bipartite graph and then run their algorithms on it. For example, FAP [7] is a representative label propagation-based fraud detection method in which given some prelabelled fraud users and items, the fraud signals can be iteratively propagated to other nodes on the built graphs. Holo [6] relies on the dense block assumption in which they view the behaviours in dense user-item blocks as fraud because fraud users usually perform more collective behaviours than normal users due to the limited cheating accounts and target items.

The methods discussed above have achieved great success on real-world e-commerce platforms. In addition to the insightful designs of these methods, we emphasize that their success is inseparable from their natural inductive property, which ensures that these methods can be used to detect newly emerging users and items directly without any additional change. This is an essential property in the fraud detection area as fraud sellers are always new and may have no relation with existing detected fraud sellers. However, in spite of their good inductive ability, the drawback of these methods is that they are all heuristic and rely on human designs, thus usually having limited performance.

Learning-based techniques for graphs, such as graph representation learning [11], [12], [13], [14] and graph neural networks(GNNs) [15], [16], [17], [18], [19], have gradually become mature in recent years, and some researchers have applied them in fraud detection [20], [21], [22], [23], [24] and achieved much better results than existing heuristic techniques. However, these methods either heavily rely on provided content features that cannot always be available in practice to obtain good performance, or need to train node ID embeddings, thus suffering a serious performance decrease in the inductive scenario because there are usually many behaviours on new users and items with untrained ID embeddings [24].

In this paper, we propose a subgraph-based method for collective fraud detection without losing the inductive property when only structural information is available, which we call SubGNN. Specifically, SubGNN is based on subgraph reasoning [25], [26]. We focus on the behaviour (edges on user-item graphs) classification task. For each candidate user-item edge, first, we extract the subgraph around it. Then, instead of using the common global ID embeddings as the node features, we mark the nodes with new label IDs on the subgraph so that the model can be entity-independent and have a better generalization ability to the edges on new users and items. Finally, we design a powerful relational graph isomorphism network (R-GIN) that has a strong expression ability and can theoretically approximate the upper bound of the Weisfeiler-Lehman (WL) graph isomorphism test method [27] on heterogeneous graphs (bipartite graphs with normal and fraudulent edges in this paper). Benefiting from it, SubGNN can learn complex knowledge reasoning rules on the relabelled heterogeneous subgraphs and perform precise fraud detection. In our experiments, we compare SubGNN with three kinds of representative methods, including label propagation, dense block mining and GNN-based solutions, on publicly available Amazon and Yelp datasets; and the results clearly show the advantages and good prospects of our method. Besides, we test SubGNN on another fraudulent purchase behaviour (transaction) dataset that we collect from a famous e-commerce platform, Taobao. The results show that SubGNN can achieve a very high precision score (0.99+) while recalling more than 90% of fraud examples.