[Paper Review] metricDTW: local distance metric learning in Dynamic Time Warping

Jiaping Zhao, Zerong Xi+1 more|arXiv (Cornell University)|Jun 11, 2016

Time Series Analysis and Forecasting18 references3 citations

TL;DR

This paper proposes metricDTW, a method that learns multiple local Mahalanobis distance metrics within and between descriptor clusters to improve k-nearest neighbor classification of temporal sequences using Dynamic Time Warping (DTW). By replacing the standard Euclidean distance with learned Mahalanobis distances along DTW alignment paths, the approach achieves state-of-the-art performance on 70 UCR time series datasets, significantly outperforming 1NN-DTW with p-values < 1.1×10⁻⁴.

ABSTRACT

We propose to learn multiple local Mahalanobis distance metrics to perform k-nearest neighbor (kNN) classification of temporal sequences. Temporal sequences are first aligned by dynamic time warping (DTW); given the alignment path, similarity between two sequences is measured by the DTW distance, which is computed as the accumulated distance between matched temporal point pairs along the alignment path. Traditionally, Euclidean metric is used for distance computation between matched pairs, which ignores the data regularities and might not be optimal for applications at hand. Here we propose to learn multiple Mahalanobis metrics, such that DTW distance becomes the sum of Mahalanobis distances. We adapt the large margin nearest neighbor (LMNN) framework to our case, and formulate multiple metric learning as a linear programming problem. Extensive sequence classification results show that our proposed multiple metrics learning approach is effective, insensitive to the preceding alignment qualities, and reaches the state-of-the-art performances on UCR time series datasets.

Motivation & Objective

• To improve kNN classification of temporal sequences by learning distance metrics that capture data regularities beyond standard Euclidean distance.
• To address the limitation of using a single, global distance metric in DTW, which may not reflect local data structures.
• To develop a local metric learning framework within DTW that adapts to subsequence patterns via descriptor clustering.
• To achieve state-of-the-art performance on UCR time series benchmark datasets using a learned metric in DTW-based 1NN classification.
• To demonstrate that metric learning is robust to the quality of preceding DTW alignments, making it practical for real-world applications.

Proposed method

• The method uses temporal sequence alignment via standard DTW to obtain alignment paths, treating them as fixed inputs.
• Each point in a sequence is represented by a descriptor (e.g., raw subsequence, HOG-1D, gradient), capturing local structural context.
• Descriptors are clustered via k-means into k groups, and a separate Mahalanobis distance metric is learned for each cluster and between every pair of clusters.
• The metric learning is formulated as a linear programming problem based on the Large Margin Nearest Neighbor (LMNN) framework, optimizing for large margins between different classes.
• The DTW distance is redefined as the sum of learned Mahalanobis distances along the alignment path, replacing the standard Euclidean distance.
• The approach is applied after alignment, making it independent of alignment quality and compatible with any existing DTW alignment method.

Experimental results

Research questions

• RQ1 Can learning multiple local Mahalanobis distance metrics improve kNN classification performance in DTW-based time series classification?
• RQ2 How sensitive is the proposed metric learning to the quality of the preceding DTW alignment?
• RQ3 Does the use of local descriptors and cluster-specific metrics lead to consistent performance gains across diverse time series datasets?
• RQ4 Can the proposed method outperform the standard 1NN-DTW classifier, which uses Euclidean distance and standard DTW alignment?
• RQ5 How do hyperparameters such as the number of descriptor clusters affect the final classification accuracy?

Key findings

• The proposed metricDTW method achieves state-of-the-art performance on 70 UCR time series datasets, significantly outperforming the standard 1NN-DTW baseline with p-values of 1.1×10⁻⁴ and 1.8×10⁻⁵ for the metric learning and no-metric variants, respectively.
• The method improves classification accuracy across all three descriptors (raw-subsequence, HOG-1D, gradient), with statistically significant gains confirmed by signed rank Wilcoxon tests (p < 0.05 for all).
• Even when the preceding DTW alignments vary in quality (as measured by alignment error), the subsequent metric learning consistently improves 1NN performance, indicating robustness to alignment errors.
• The number of descriptor clusters (k ∈ {5,10,15,20,25,30}) has a significant positive effect on performance, with all k values yielding statistically significant improvements (p < 0.05) over the Euclidean baseline.
• The method achieves consistent and substantial performance gains regardless of the descriptor used for alignment, demonstrating the generalizability of the local metric learning approach.
• Iterative joint alignment and metric learning was found to degrade performance, suggesting that decoupling alignment and metric learning is preferable for current formulation.