Research Contributions

With a background in academia, the team at Flex Analytics has a strong commitment to contributing to the research community. We believe that open-source software and research are essential for technological advancement.

Below is a selection of our most notable research contributions:

Machine Learning

• Do Not Sleep on Traditional Machine Learning: Simple and Interpretable Techniques Are Competitive to Deep Learning for Sleep Scoring: Demonstrates how traditional machine learning methods can match or even outperform deep learning for sleep stage classification, offering greater interpretability and thereby making it more suitable for clinical adoption.

• Left-Right Swapping and Upper-Lower Limb Pairing for Robust Multi-Wearable Workout Activity Detection: Introduces two novel multi-wearable data augmentation techniques; left-right swapping and upper-lower limb pairing, to overcome inconsistencies in wearable orientation and improve model robustness.
  🏆 1st place in the 2024 WEAR challenge

• Magnitude and Rotation Invariant Detection of Transportation Modes with Missing Data Modalities: Presents a novel rotation-invariant feature aggregation technique, validated on multimodal transportation mode detection using smartphone data, that improves model robustness and accuracy in real-world scenarios.
  🏆 1st place in the 2024 SHL challenge

Time Series Analysis

• tsflex: Flexible Time Series Processing & Feature Extraction: A Python library for high-performance, memory-efficient time series processing and feature extraction that seamlessly handles irregular, asynchronous data, multiple window-stride setups, and integrates with existing tools.

Time Series Visualization

• Plotly-Resampler: Effective Visual Analytics for Large Time Series: A Python library that facilitates efficient visualization and analysis of time series data, enabling the design and development of novel line chart aggregation techniques while also being production-ready for large-scale applications and integrating seamlessly into standard data science workflows.

• MinMaxLTTB: Leveraging MinMax-Preselection to Scale LTTB: Introduces a novel algorithm that improves the scalability of the Largest-Triangle-Three-Buckets (LTTB) downsampling by utilizing MinMax prefetching, achieving similar visual quality with over 10x faster performance and supporting multi-threaded execution.

• tsdownsample: Efficient Downsampling Techniques for Large Time Series: A Python library that provides extremely fast, memory-efficient time series downsampling techniques, including MinMaxLTTB, that are optimized in Rust by leveraging CPU-specific SIMD instructions and multithreading, enabling interactive data visualization of large datasets.