Summary
This paper introduces the GAN-RL method for event outlier detection in continuous time event streams. The method proceeds by learning an encoder for event streams based on an LSTM architecture, which outputs a latent vector for each time point. Then, RL agents are trained to tamper with the TPP by adding and removing point occurrences (thus "generating" a new sequence). Finally, a discriminator is simultaneously trained to identify artificially generated samples and acts as the reward function of the RL agents.
At test time, the RL agents treat each sequence as an episode and score each point and interval as outliers. AUROC of ground truth outliers (omissions and commissions also labelled randomly by authors) are reported.
Overall the paper is well written, sets up and explores the continuous time event outlier detection problem well. However the paper is missing a large ablation study and empirical results to justify the amount of complexity inherent in the methodology is missing.
Strengths
The paper proposes a strong method for event outlier detection in continuous time. Its combination of different ideas from deep reinforcement learning and neural sequence models is novel and interesting, and may be widely applicable in other tasks. It would be interesting to try the ideas in the paper, for example, in time series anomaly detection.
The paper also makes a significant effort to rigorously set up the event outlier detection task where an outlier / anomaly may simply be that an event is "missing." Its treatment of the topic and its survey of related works is insightful and accurate.
Weaknesses
The main weakness of the paper is its lack of ablation studies and reproducibility. The paper introduces an overly complex architecture for a task where naive baselines are often available in practice. The architecture comprises an LSTM for encoding, multiple RL agents (where the policy of one depends on the policy of the other--and not vice versa), a discriminator architecture that could also be seen as the "reward function," design choices in the neural networks in the form of "continuous time" LSTMs, spectral normalization and attention. Only one of these design choices is subjected to an ablation study, where it is seen suprisingly that it is is critical to the success of the method. This invites the question how other complex architectural additions benefit the method and whether they are necessary. In the absence of such a study + open-source code to easily reproduce the work, the proposed method does not appear practical.
Similarly, the paper only compares with the PPOD baseline. However many other naive baselines are missing, such as simply learning a simple TPP model using a PPOD like procedure. Even simpler, see the RND and LEN baselines in the PPOD paper, or baselines in the Shchur et al paper applied on a discretized space, etc. i.e., there appear to be a continuum of ideas from naive baselines to the complex architecture of GAN-RL that need to be explored especially seeing that the task definition is not clear and there is no common task / benchmark for event outlier detection. (i.e., the authors define the task as well as solve it.)
Questions
1. I do not follow the claim that the method is able to "learn from contaminated data" where all other methods assume clean data. I would beg to differ, as all unsupervised outlier detection methods would in practice learn from contaminated data and demonstrate empirically that they are resilient to this contamination. Could you explain this further with references to where else the empirical success of outlier detection methods is conditioned on the availability of clean (i.e., one-class) data and do not work with contaminated (i.e., "two-class," labels not available) data?
Rating
5: marginally below the acceptance threshold
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.