Barely Supervised Learning for Graph-Based Fraud Detection

Authors: Hang Yu, Zhengyang Liu, Xiangfeng Luo

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conducted comparative experiments on five real-world datasets using ten currently representative or novel methods to test the performance of our proposed approach.
Researcher Affiliation Academia Hang Yu*, Zhengyang Liu*, Xiangfeng Luo School of Computer Engineering and Science, Shanghai University, Shanghai, China {yuhang, zhengyangliu, luoxf}@shu.edu.cn
Pseudocode No The paper describes the model components and equations but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions that code for other compared methods is published, but does not state that the source code for their own method (BSL) is openly available or provide a link.
Open Datasets Yes We employ five datasets to evaluate the effectiveness of BSL, with their specific statistical characteristics presented in Table I. Amazon and Yelp (Mc Auley and Leskovec 2013) are datasets of fraudulent comments... Elliptic (Weber et al. 2019) constructs a transaction network of bitcoin...
Dataset Splits Yes In the experiment, each dataset is divided into three parts a small proportion as labeled samples, another part as unlabeled samples for barely supervised learning, and the remainder used as the test set. For Amazon, Yelp, Elliptic, and T-Finance datasets, the split is 1% : 10%: 89%, while that for T-Social is 0.01%: 0.1%: 99.89% due to its large scale.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments.
Software Dependencies Yes Our method is implemented in Py Torch 1.13.1 (Paszke et al. 2019) and Py G (Fey and Lenssen 2019) with Python 3.9, and DGL (Wang et al. 2019c) is used for preprocessing the data.
Experiment Setup Yes The parameters are optimized by Adam (Kingma and Ba 2014), while the learning rate is set as 0.01, and the weight decay rate is 1e-5. We set the embedding size to 96, and the batch size to 256 on five datasets.