TsKAN: An Transparent Architecture for Improving the Interpretability of Multivariate Time Series Forecasting

In recent years, numerous deep learning models have been proposed for Multi-variate Time Series (MTS) forecasting, with Transformer-based models showing significant potential due to their ability to capture long-term dependencies. However, existing models based on MLPs or Transformers often suffer from a lack of interpretability due to their large parameter sizes, which can be problematic in many real-world applications. To address this issue, we propose TimeKAN, a model based on Kolmogorov-Arnold Networks. The KAN model offers two key advantages: (1) it achieves accuracy comparable to MLPs with significantly fewer parameters, and (2) its parameters can be symbolized, which makes it possible to interpret the meaning of the parameters. Additionally, instead of the usual attention mechanisms, we designed a Multi-Scale Patching (MSP) module for MTS that allows for more flexible and simple multi-patching and effectively extracts both temporal and cross-dimensional features. By leveraging this strategy along with KAN, TimeKAN constructs a hierarchical structure capable of utilizing information across different scales, leading to highly accurate predictions. Extensive experiments on six real-world datasets demonstrate that TimeKAN outperforms state-of-the-art (SOTA) methods in terms of predictive performance. Furthermore, we interpret TimeKAN by visualizing its learning process for extracting symbolized features, opening the black box and revealing meaningful patterns within the time series.