Source-linked AI summary
TSExplorer: An interactive data annotation and exploration tool for time-series data
Einari Vaaras, Manu Airaksinen, Okko Räsänen
TL;DR
Time-series analysis workflows often underuse high-dimensional feature representations and rely on sequential annotation, static visualizations, or aggregate statistics. TSExplorer addresses this gap with a cross-platform GUI that links interactive 2D visualizations to multimodal sample views and flexible annotation workflows. It supports exploration and annotation across unlabeled, partially-labeled, and fully-labeled datasets, with resource demands shaped by preprocessing and data-loading choices.
Problem
Existing time-series annotation and analysis workflows often discard high-dimensional feature representations or make limited use of their structure.
Method
TSExplorer provides interactive 2D visualizations, linked sample views, feature-representation switching, algorithmic sample selection, labeling controls, and customizable extensible widgets.
Results
TSExplorer supports exploratory analysis and annotation for unlabeled, partially-labeled, and fully-labeled datasets, including outlier identification, incremental annotation, feature-space inspection, and label revision.
Takeaways & Limitations
The tool provides a flexible research interface for inspecting time-series samples alongside multiple complementary views of their high-dimensional representations.
Takeaways & Limitations
Large datasets can require substantial memory and computation, so precomputing 2D visualizations and loading data on demand are recommended to reduce overhead.
Abstract
from arXiv · showhide
We present TSExplorer, a cross-platform tool for interactive annotation and exploration of time-series data. The tool enables users to inspect high-dimensional datasets through multiple complementary 2D visualizations derived from high-dimensional feature representations. TSExplorer is designed as a general-purpose research tool supporting a wide range of workflows, including exploratory data analysis, annotation of unlabeled or partially-labeled datasets, comparison of feature representations, and post-hoc inspection and refinement of existing labels with interactive visual feedback.
1. Introduction
TSExplorer addresses limited use of high-dimensional time-series representations by providing a flexible GUI for interactive annotation, analysis, and interpretation. It supports exploration through interactive 2D visualizations rather than relying only on sequential annotation, static visualizations, or aggregate statistics.
- Motivation: High-dimensional representations for speech, video, physiological signals, sensor recordings, and multimodal data are often discarded or only partially exploited in annotation and analysis workflows.Existing workflows frequently use sequential annotation, static visualizations, or aggregate statistics such as feature variances, class-wise averages, and performance metrics.
- Contribution: TSExplorer is a general-purpose GUI for interactive time-series data annotation, analysis, and interpretation.The tool is designed to support flexible workflows rather than a single predefined procedure.
- Contribution: The tool enables users to explore high-dimensional datasets through interactive 2D visualizations.TSExplorer is publicly available on GitHub and had been previously described and evaluated according to the supplied passage.
2. Interactive Data Annotation and Exploration with TSExplorer
TSExplorer combines scatter-based dataset visualization with linked sample views, multiple dimensionality-reduction methods, flexible selection and labeling, and customizable widgets. These capabilities support annotation and exploration across unlabeled, partially-labeled, and fully-labeled datasets.
- Interactive visualization: The GUI displays the dataset as a 2D scatter plot and presents linked sample views when users select data points.Linked views can include audio, video, and signal waveforms.
- Interactive visualization: Users can switch among t-SNE, PCA, and UMAP visualizations and among high-dimensional feature representations within or across modalities.These 2D visualizations highlight different aspects of the underlying high-dimensional data.
- Annotation workflow: Samples can be selected manually or with Random, Ordered, and Farthest-first strategies, then queued for later inspection and annotation.The scatter plot also supports zooming, and labels can be modified through a drop-down menu or keyboard shortcuts.
- Annotation workflow: TSExplorer uses sample colors to reflect labels and a large ring to identify the currently selected sample.Unlabeled samples are green by default and the selected sample is yellow by default.
- Supported workflows: The tool supports unlabeled, partially-labeled, and fully-labeled datasets for exploration, incremental annotation, separability and consistency evaluation, and label revision.The supported workflows include identifying outliers, examining label distributions, comparing feature representations, and quality assurance.
- Extensibility: Five widget types are provided, layouts and widget settings are customizable, and new widgets, 2DV methods, and sample-selection strategies can be added.The existing components are designed to be modified in a straightforward manner.
3. System Design and Implementation
TSExplorer uses a cross-platform Python implementation in which feature representations are computed offline and 2D visualizations may be computed offline or during use. Resource demands depend on dataset size, feature configuration, and whether data are preloaded or loaded on demand.
- Implementation: TSExplorer runs on Windows, Linux, and macOS using PySide6 for the GUI, PyQtGraph for visualization, and VLC Media Player for audio and video playback.The system structure is illustrated in Figure 2.
- Computational design: Feature representations are computed offline, while 2D visualization computation is the most demanding process during use.The tool primarily performs file input/output during interactive use and saves computed 2D visualizations to avoid repeated computation.
- Computational design: 1.8 GB of RAM and 8 minutes on one CPU core are required to compute a t-SNE embedding for 100,000 160-dimensional samples.Precomputing 2D visualizations is recommended for large datasets to reduce computational overhead.
- Memory requirements: 5.1 GB of RAM is used by a 100,000-sample speech dataset with preloaded audio, three feature types, and their corresponding 2D visualizations.Loading audio as individual files on disk and on demand reduces RAM usage below 0.4 GB; with no data loaded, the tool uses approximately 0.3 GB.
5. Generative AI Use Disclosure
Generative AI was used only in a limited programming-support role.
- Generative AI supported programming through syntax guidance and debugging.