KAN-AD: Time Series Anomaly Detection with Kolmogorov–Arnold Networks

Time series anomaly detection (TSAD) underpins real-time monitoring in cloud services and web systems, allowing rapid identification of anomalies to prevent costly failures. Most TSAD methods driven by forecasting models tend to overfit by emphasizing minor fluctuations. Our analysis reveals that effective TSAD should focus on modeling "normal" behavior through smooth local patterns. To achieve this, we reformulate time series modeling as approximating the series with smooth univariate functions. The local smoothness of each univariate function ensures that the fitted time series remains resilient against local disturbances. However, a direct KAN implementation proves susceptible to these disturbances due to the inherently localized characteristics of B-spline functions. We thus propose KAN-AD, replacing B-splines with truncated Fourier expansions and introducing a novel lightweight learning mechanism that emphasizes global patterns while staying robust to local disturbances. On four popular TSAD benchmarks, KAN-AD achieves an average 15% improvement in detection accuracy (with peaks exceeding 27%) over state-of-the-art baselines. Remarkably, it requires fewer than 1,000 trainable parameters, resulting in a 50% faster inference speed compared to the original KAN, demonstrating the approach's efficiency and practical viability.

Spotting unusual activity (anomalies) in the continuous data streams from online services and cloud systems is crucial. If we miss these blips, it can lead to costly failures. However, many current methods get easily sidetracked by normal, minor fluctuations in the data. They try to predict every little wiggle, which means they often struggle to get a clear picture of what "normal" truly looks like and can miss the real warning signs.We realized that to reliably detect anomalies, a system needs to understand the smooth, underlying patterns of normal behavior, rather than getting caught up in tiny details. So, we developed a new technique called KAN-AD. It learns to represent time series data using smooth mathematical functions. Specifically, KAN-AD uses a method based on Fourier expansions—a way to describe complex patterns as a sum of simple waves—to focus on these stable, global trends, making it robust against misleading local disturbances or "noise."Our KAN-AD approach significantly improves the ability to catch real problems. In tests on standard benchmarks, it boosted anomaly detection accuracy by an average of 15%, with improvements sometimes exceeding 27%, compared to other leading methods. Remarkably, KAN-AD achieves this with very few adjustable parts (less than 1,000 parameters) and is 50% faster than the original advanced technique it builds upon. This makes it a highly efficient and practical tool for keeping critical online systems running smoothly and safely.