Detecting problematic transactions in a consumer-to-consumer e-commerce network

Abstract

Providers of online marketplaces are constantly combatting against problematic transactions, such as selling illegal items and posting fictive items, exercised by some of their users. A typical approach to detect fraud activity has been to analyze registered user profiles, user’s behavior, and texts attached to individual transactions and the user. However, this traditional approach may be limited because malicious users can easily conceal their information. Given this background, network indices have been exploited for detecting frauds in various online transaction platforms. In the present study, we analyzed networks of users of an online consumer-to-consumer marketplace in which a seller and the corresponding buyer of a transaction are connected by a directed edge. We constructed egocentric networks of each of several hundreds of fraudulent users and those of a similar number of normal users. We calculated eight local network indices based on up to connectivity between the neighbors of the focal node. Based on the present descriptive analysis of these network indices, we fed twelve features that we constructed from the eight network indices to random forest classifiers with the aim of distinguishing between normal users and fraudulent users engaged in each one of the four types of problematic transactions. We found that the classifier accurately distinguished the fraudulent users from normal users and that the classification performance did not depend on the type of problematic transaction.

Temporal Motifs for Financial Networks: A Study on Mercari, JPMC, and Venmo Platforms

“…Li et al (2020) proposed a metric to measure the high volume of money flows and spot money laundering behavior. Kodate et al (2020) measured the local network characteristics, such as the local degree and clustering coefficient, and used random forest classifiers to detect problematic transactions in online e-commerce networks. Note that most of the existing studies do not consider the timestamp and chronological order of the events, which may be crucial in detecting the fraudulent activities.…”

“…For both datasets, we utilize temporal motifs to detect fraudulent users. We extract temporal motif features from the egocentric networks and compare them against a baseline model that only considers simple graph features (Kodate et al 2020).…”

“…In a related study (Kodate et al 2020), the authors trained fraud detection classifiers by using simple graph features. They considered the following features for each node:…”

“…We create a feature vector for each user based on the number of temporal motifs containing the ego user and those do not. We use four different classifiers on these features, logistic regression, support vector machine, random forest, and XGBoost, and compare them against a robust baseline that utilizes simple graph features, such as degree and number of events (Kodate et al 2020). We show that temporal motif based features consistently outperforms the baseline, reaching up to 0.93 and 0.81 F1 scores on Mercari and JPMC networks, respectively.…”

BiRank vs PageRank: Using SNA on Company Register Data for Fiscal Risk Prediction

“…A wide range of SNA methods that are (or can be) used for fraud prediction exists in the literature. Network visualization and ego-centric analysis (and quite similarly the snowball method) can be used to identify links with known arXiv:2209.01945v1 [cs.SI] 5 Sep 2022 fraudulent entities (for example in [2]). Collective inference or collective classification methods, such as Gibbs sampling or loopy believe propagation, are commonly used in [3].…”