tpcp: Tiny Pipelines for Complex Problems - A set of framework independent helpers for algorithms development and evaluation

Abstract

During algorithm development and analysis researchers regularly use software libraries developed for their specific domain. With such libraries, complex analysis tasks can often be reduced to a couple of lines of code. This not only reduces the amount of implementation required but also prevents errors.The best developer experience is usually achieved when the entire analysis can be represented with the tools provided by a single library. For example, when an entire machine learning pipeline is represented by a scikit-learn pipeline (Pedregosa et al., 2018), it is extremely easy to switch out and train algorithms. Furthermore, train/test leaks and other methodological errors at various stages in the analysis are automatically prevented -even if the user might not be aware of these issue.However, if the performed analysis gets too complex, too specific to an application domain, or requires the use of tooling and algorithms from multiple frameworks, developers lose a lot of the benefits provided by individual libraries. In turn, the required skill level and the chance of methodological errors rise.With tpcp we attempt to overcome the issue by providing higher-level tooling and structure for algorithm development and evaluation that is independent of the frameworks required for the algorithm implementation.

Gaitmap—An Open Ecosystem for IMU-Based Human Gait Analysis and Algorithm Benchmarking

“…The presented ecosystem consists of four core packages that all build on top the domain-agnostic tpcp package [33] (Fig. 1).…”

“…Ideally, this results in a larger user base of the more general packages, increasing the chance of building stable and sustainable communities around them. For example, the tpcp package [33], which was developed as the foundation for all the other packages, has already been used in the fields of digital psychology [59] and is being used in further applications fields our research group.…”

DISPEL: A Python Framework for Developing Measures From Digital Health Technologies

“…Available open-source libraries for the development of SDMs, in fact, are often limited to very specific applications or device sensor locations [4] , [5] , [6] . Even recent efforts towards a code library for integration of diverse biomarkers [7] or for effective standardization [8] , [9] do not seem to have yet achieved the needed level of traceability and generalizability.…”

Data-Driven Radio Propagation Modeling using Graph Neural Networks

“…To overcome these hurdles, open-source tools with validated algorithms have been deployed by the community, but often for very specific applications [1][2][3]. Recent remarkable efforts towards a solution for integration of diverse biomarkers [4] or for effective standardization [5,6] do not yet seem to have achieved the needed level of traceability and generalizability. These stumbling blocks hinder SDM adoption, especially when faced with the complexity of datasets collected within clinical trials.…”