TL;DR
CaPulse is a causality-based framework that detects anomalies in time series by tuning into their underlying causal rhythms, effectively addressing data challenges and outperforming existing methods.
Contribution
Introduces CaPulse, a novel causality-driven approach utilizing structural causal models and periodic normalizing flows for improved anomaly detection in time series.
Findings
Outperforms existing methods with 3-17% AUROC improvements
Enhances interpretability of anomaly detection
Effective on seven real-world datasets
Abstract
Time series anomaly detection has garnered considerable attention across diverse domains. While existing methods often fail to capture the underlying mechanisms behind anomaly generation in time series data. In addition, time series anomaly detection often faces several data-related inherent challenges, i.e., label scarcity, data imbalance, and complex multi-periodicity. In this paper, we leverage causal tools and introduce a new causality-based framework, CaPulse, which tunes in to the underlying causal pulse of time series data to effectively detect anomalies. Concretely, we begin by building a structural causal model to decipher the generation processes behind anomalies. To tackle the challenges posed by the data, we propose Periodical Normalizing Flows with a novel mask mechanism and carefully designed periodical learners, creating a periodicity-aware, density-based anomaly…
Peer Reviews
Decision·Submitted to ICLR 2025
The paper makes an effort to integrate multiple algorithmic building blocks into a learning system.
It seems that the learning problem could benefit from a clearer explanation. The objective function appears to be conditioned by $ C_\text{ind} $ and $C_0$, but the criteria for selecting these values are not immediately obvious. This aspect is essential, especially when considering unsupervised learning tasks with latent variables. It is possible that a two-stage optimization strategy, where $C$ parameters and other model parameters are optimized in an alternating fashion, has been implemented,
The paper's main strength is its integration of causal inference with time series anomaly detection and proposed solution for handling multiple periodicities. This method offers some degree of interpretability through SHAP. The authors do establish strong theoretical foundation for their method.
In my view the main weakness is with the empirical evaluation. The proposed method is extremely complex, likely computationally demanding with a large number of hyperparameters. The authors do perform some sensitivity analysis but it is limited. The baselines used for the empirical comparison relies exclusively on similarly complex baselines. The authors do not compare to simple, algorithmic baselines or methods such as Matrix Profile which have proven to outperform state of the art at a computa
S1. Time series anomaly detection is important to various domains. S2. There are quite a few nice illustrations. S3. This work focuses on an important problem that could have real-world applications. S4. The figures and tables used in this work are clear and easy to read.
W1. The paper conducts ablation experiments solely on two datasets, as shown in Table 2. This narrow focus raises concerns about the generalizability of the findings. A more comprehensive analysis involving additional datasets could provide valuable insights into the method's performance and limitations. W2. The approach presented appears to lack novelty, as it primarily builds upon established methods of causal inference and frequency domain analysis without offering significant advancements.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
