ALLIE: Active Learning on Large-scale Imbalanced Graphs

被引:9
作者
Cui, Limeng [1 ]
Tang, Xianfeng [2 ]
Katariya, Sumeet [2 ]
Rao, Nikhil [2 ]
Agrawal, Pallav [2 ]
Subbian, Karthik [2 ]
Lee, Dongwon [1 ]
机构
[1] Penn State Univ, Philadelphia, PA 16801 USA
[2] Amazon, Seattle, WA USA
来源
PROCEEDINGS OF THE ACM WEB CONFERENCE 2022 (WWW'22) | 2022年
基金
美国国家科学基金会;
关键词
Graph neural networks; fraud detection; active learning; reinforcement learning; REDUCTION;
D O I
10.1145/3485447.3512229
中图分类号
TP3 [计算技术、计算机技术];
学科分类号
0812 ;
摘要
Human labeling is time-consuming and costly. This problem is further exacerbated in extremely imbalanced class label scenarios, such as detecting fraudsters in online websites. Active learning selects the most relevant example for human labelers to improve the model performance at a lower cost. However, existing methods for active learning for graph data often assumes that both data and label distributions are balanced. These assumptions fail in extreme rare-class classification scenarios, such as classifying abusive reviews in an e-commerce website. We propose a novel framework ALLIE to address this challenge of active learning in large-scale imbalanced graph data. In our approach, we efficiently sample from both majority and minority classes using a reinforcement learning agent with imbalance-aware reward function. We employ focal loss in the node classification model in order to focus more on rare class and improve the accuracy of the downstream model. Finally, we use a graph coarsening strategy to reduce the search space of the reinforcement learning agent. We conduct extensive experiments on benchmark graph datasets and real-world e-commerce datasets. ALLIE out-performs state-of-the-art graph-based active learning methods significantly, with up to 10% improvement of F1 score for the positive class. We also validate ALLIE on a proprietary e-commerce graph data by tasking it to detect abuse. Our coarsening strategy reduces the computational time by up to 38% in both proprietary and public datasets.
引用
收藏
页码:690 / 698
页数:9
相关论文
共 54 条
  • [1] Adhikari B, 2021, AAAI CONF ARTIF INTE, V35, P15232
  • [2] Aggarwal CC, 2014, CH CRC DATA MIN KNOW, P457
  • [3] Aggarwal U, 2020, IEEE WINT CONF APPL, P1417, DOI 10.1109/WACV45572.2020.9093475
  • [4] Bachman P, 2017, PR MACH LEARN RES, V70
  • [5] Geometric Deep Learning Going beyond Euclidean data
    Bronstein, Michael M.
    Bruna, Joan
    LeCun, Yann
    Szlam, Arthur
    Vandergheynst, Pierre
    [J]. IEEE SIGNAL PROCESSING MAGAZINE, 2017, 34 (04) : 18 - 42
  • [6] A systematic study of the class imbalance problem in convolutional neural networks
    Buda, Mateusz
    Maki, Atsuto
    Mazurowski, Maciej A.
    [J]. NEURAL NETWORKS, 2018, 106 : 249 - 259
  • [7] Cai HY, 2017, Arxiv, DOI arXiv:1705.05085
  • [8] Towards Kron reduction of generalized electrical networks
    Caliskan, Sina Yamac
    Tabuada, Paulo
    [J]. AUTOMATICA, 2014, 50 (10) : 2586 - 2590
  • [9] SMOTE: Synthetic minority over-sampling technique
    Chawla, Nitesh V.
    Bowyer, Kevin W.
    Hall, Lawrence O.
    Kegelmeyer, W. Philip
    [J]. 2002, American Association for Artificial Intelligence (16)
  • [10] Chen X, 2019, PROCEEDINGS OF THE TWENTY-EIGHTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE, P2123