Con4m: Context-aware Consistency Learning Framework for Segmented Time Series Classification

TL;DR

Con4m tackles segmented time series classification under multiple duration classes by explicitly modeling contextual information across consecutive segments and harmonizing boundary label inconsistencies. It introduces a three-component framework: a Gaussian-prior continuous contextual encoder, a context-aware predictor with neighbor-consistency and monotonicity constraints, and a label-consistency training scheme with curriculum-based harmonization. The approach yields robust improvements across four MVD datasets, demonstrating resilience to boundary disturbances and providing evidence that contextual data and label coherence substantially boost discriminative power in segmented TSC. The work highlights the importance of temporal dependencies in segmented time series and offers practical, end-to-end mechanisms to leverage them in real-world applications.

Abstract

Time Series Classification (TSC) encompasses two settings: classifying entire sequences or classifying segmented subsequences. The raw time series for segmented TSC usually contain Multiple classes with Varying Duration of each class (MVD). Therefore, the characteristics of MVD pose unique challenges for segmented TSC, yet have been largely overlooked by existing works. Specifically, there exists a natural temporal dependency between consecutive instances (segments) to be classified within MVD. However, mainstream TSC models rely on the assumption of independent and identically distributed (i.i.d.), focusing on independently modeling each segment. Additionally, annotators with varying expertise may provide inconsistent boundary labels, leading to unstable performance of noise-free TSC models. To address these challenges, we first formally demonstrate that valuable contextual information enhances the discriminative power of classification instances. Leveraging the contextual priors of MVD at both the data and label levels, we propose a novel consistency learning framework Con4m, which effectively utilizes contextual information more conducive to discriminating consecutive segments in segmented TSC tasks, while harmonizing inconsistent boundary labels for training. Extensive experiments across multiple datasets validate the effectiveness of Con4m in handling segmented TSC tasks on MVD. The source code is available at https://github.com/MrNobodyCali/Con4m.

Figures (7)

• Figure 1: (a) Reasonable model predictions exhibit coherence across consecutive segments rather than repeated interruptions. (b) In the healthcare domain, different physicians have varying annotations regarding the start and end times of seizure waves. (c) Based on the proximity to the boundary, we divide each class sequence into 5 levels, from which an equal number of segments are sampled. A one-layer MLP is trained on the segments from each level respectively for the same number of epochs. (d) We visualize the predicted probability of the trained MLP for each level. We observe that as the segments approach the boundaries, the model finds it increasingly challenging to make correct classifications, resulting in more extreme wrong predictions. This strongly underscores the significance of handling boundary segments.
• Figure 2: Overview of Con4m. (a) Overview of continuous contextual representation encoder in Con4m. The leftmost part shows the details of Con-Attention. The right part of the figure shows the architecture of Con-Transformer and the whole encoder of Con4m. (b) Overview of context-aware coherent class prediction and consistent label training framework in Con4m. The right part describes the neighbor class consistency discrimination task and the prediction behavior constraint. The leftmost part presents the training and inference details for label harmonization.
• Figure 3: Comparison results of symmetric disturbance and label substitution experiments.
• Figure 4: Case study for a continuous time interval in SEEG testing set. The C-score, introduced by the ClaSP model 10.1007/s10618-023-00923-x, assessing the ability of models to recognize segmentation boundaries by measuring the trade-off between precision (correctly identified change points) and recall (finding all true change points).
• Figure 5: Cases for Tanh fitting.

Theorems & Definitions (3)

• Theorem 2.1
• proof
• Definition 3.1