Problem
A fixed document–token graph can classify documents seen during construction, yet struggle to incorporate new text. This limits the usefulness of an otherwise strong offline text classifier in a streaming system.
Approach
ContGCN updates the document–token graph for each batch during training and inference. An occurrence memory and a self-supervised contrastive objective support updates without requiring labels for each arriving document.
Evidence
The paper evaluates five public datasets and reports an 8.86% performance gain in a three-month A/B test on Huawei's public-opinion analysis system. This is the paper's task-performance result, not a revenue or conversion-rate claim.
Connection
The method illustrates learning over changing graph evidence. Later recommendation-graph work addresses different data and objectives, so this connection should not be read as a shared algorithm or a direct transfer result.
Sources
Bibliographic details
- Authors
- Tiandeng Wu, Qijiong Liu, Yi Cao, Yao Huang, Xiao-Ming Wu, Jiandong Ding
- Venue
- AAAI 2023
- Status
- Published
- First published
- 2023-06-26