Abstract:Multivariate time series classification is of great importance in practical applications and is a challenging task. However, deep neural network models such as Transformers exhibit high accuracy in multivariate time series classification but lack interpretability and fail to provide insights into the decision-making process. On the other hand, traditional approaches based on decision tree classifiers offer clear decision processes but relatively lower accuracy. Swin Transformer (ST) addresses these issues by leveraging self-attention mechanisms to capture both fine-grained local patterns and global patterns. It can also model multi-scale feature representation learning, thereby providing a more comprehensive representation of time series features. To tackle the aforementioned challenges, we propose ST-Tree with interpretability for multivariate time series classification. Specifically, the ST-Tree model combines ST as the backbone network with an additional neural tree model. This integration allows us to fully leverage the advantages of ST in learning time series context while providing interpretable decision processes through the neural tree. This enables researchers to gain clear insights into the model's decision-making process and extract meaningful interpretations. Through experimental evaluations on 10 UEA datasets, we demonstrate that the ST-Tree model improves accuracy in multivariate time series classification tasks and provides interpretability through visualizing the decision-making process across different datasets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to improve the interpretability of the model while maintaining high accuracy in multivariate time - series classification tasks. Specifically, existing deep neural network models such as Transformer exhibit high precision in multivariate time - series classification, but lack interpretability and cannot provide insights into the decision - making process; while traditional decision - tree - based classifiers, although providing a clear decision - making process, have relatively low accuracy. Therefore, the paper proposes a method that combines the advantages of both - ST - Tree, aiming to achieve a balance between high accuracy and interpretability by combining Swin Transformer (ST) and the neural tree model. ### Main contributions of the paper: 1. **Propose the ST - Tree model**: This model combines Swin Transformer (ST) as a feature extractor and time - slice generator, and a neural tree module to provide an interpretable decision - making process. 2. **Transparent decision - making process**: The decision - making process of the neural tree model itself is transparent, and the judgment criteria of each node can be clearly understood through the prototype solutions of the nodes. 3. **Experimental verification**: Through experimental evaluations on 10 UEA datasets, it is proven that the ST - Tree model has improved accuracy in multivariate time - series classification tasks and has good interpretability. 4. **Visualization**: Visualize the tree model on different types of databases, demonstrating the good interpretability of the model through the decision - making process. ### Method overview: 1. **Time - slice module**: Use Swin Transformer (ST) to divide the multivariate time - series into time - slices and extract features. 2. **Neural tree module**: Input the generated time - slices into the neural tree module, perform hierarchical path selection through a binary - tree structure, and finally perform label prediction at the leaf nodes. 3. **Routing score calculation**: Determine the routing direction by calculating the similarity between the time - slice and the prototype, thereby determining the branch direction. 4. **Attention mechanism**: Utilize spatial and channel attention mechanisms to refine and extract features, enhancing the representational ability of the model. ### Formula summary: - **Self - attention calculation formula**: \[ X_{\text{SA}}=\text{softmax}\left(\frac{\text{attention}(QX)\cdot(\text{attention}(KX))^T}{\sqrt{d_k}}\right)\cdot\text{attention}(VX) \] - **Similarity calculation formula**: \[ N(z_i)=\log\left(1 + \frac{1}{\max\left(\|\tilde{z}-P_i\|^2_2+\epsilon\right)}\right) \] - **Routing score calculation formula**: \[ R_{i,j}(z_i)=\begin{cases} [N(z_i)]^{1}_{0}&\text{if }j = 2\times i\\ 1 - [N(z_i)]^{1}_{0}&\text{if }j = 2\times i+ 1 \end{cases} \] - **Cumulative routing score calculation formula**: \[ \rho(z_i)=\text{Aggregate}(\{R_{i,j}(E_{i,j}(z_i))\mid j\in\text{children}(i)\}) \] Through these methods and formulas, the ST - Tree model can provide a clear decision - making process while maintaining high accuracy, thus solving the interpretability problem in multivariate time - series classification.

ST-Tree with Interpretability for Multivariate Time Series Classification

TV-Net: Temporal-Variable feature harmonizing Network for multivariate time series classification and interpretation

FormerTime: Hierarchical Multi-Scale Representations for Multivariate Time Series Classification

Multivariate Time Series Classification with Crucial Timestamps Guidance

A Deep Multi-Task Representation Learning Method for Time Series Classification and Retrieval.

Learn Decision Trees with Deep Visual Primitives

Expressing Multivariate Time Series as Graphs with Time Series Attention Transformer

An Aggregated Convolutional Transformer Based on Slices and Channels for Multivariate Time Series Classification

ViTree: Single-path Neural Tree for Step-wise Interpretable Fine-grained Visual Categorization

VSFormer: Value and Shape-Aware Transformer with Prior-Enhanced Self-Attention for Multivariate Time Series Classification

Leveraging 2D Information for Long-term Time Series Forecasting with Vanilla Transformers

SVP-T: A Shape-Level Variable-Position Transformer for Multivariate Time Series Classification

Gated Transformer Networks for Multivariate Time Series Classification

UniTST: Effectively Modeling Inter-Series and Intra-Series Dependencies for Multivariate Time Series Forecasting

Scalable Transformer for High Dimensional Multivariate Time Series Forecasting

Multiscale spatial‐temporal transformer with consistency representation learning for multivariate time series classification

Spatial-Temporal Convolutional Transformer Network for Multivariate Time Series Forecasting

MTS2Graph: Interpretable Multivariate Time Series Classification with Temporal Evolving Graphs

Densely Knowledge-Aware Network for Multivariate Time Series Classification

Enhancing Transformer Efficiency for Multivariate Time Series Classification

Ister: Inverted Seasonal-Trend Decomposition Transformer for Explainable Multivariate Time Series Forecasting