Reproduce “Unsupervised Event Outlier Detection in Continuous Time” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Somjit Nath, Yik Chau Lui, Siqi Liu
Researchers from McGill University, Mila, and Borealis AI developed GAN-RL, an unsupervised framework that leverages Generative Adversarial Networks and Reinforcement Learning to detect individual abnormal events within continuous-time event sequences. The model consistently achieved higher AUROC scores on synthetic and real-world datasets, outperforming state-of-the-art semi-supervised baselines by learning directly from unlabeled, potentially corrupted data.
Key insights
- The adversarial training loop between the RL-based Generator and the Discriminator effectively enables unsupervised learning of event normalcy and outlier detection without explicit labels.
- The use of continuous-time LSTM (cLSTM) with an attention mechanism in the Generator's encoder is crucial for capturing temporal dependencies and making informed decisions about individual events.
- The framework demonstrates robustness, performing well even when the input dataset contains a substantial proportion of corrupted (outlier-containing) sequences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.