This repository contains the open-source crossai library, which includes functionalities capable of developing
end-to-end Artificial Intelligence (AI) pipelines for uni- or multi-axis Time Series analysis. The library specifically
incorporates components in the likes of:
- Signal engineering: signal transformation, feature engineering, data augmentation
- State-of-the-Art Neural Network models, capable of training with time series data
- Explainable AI (XAI) modeling for evaluating the AI time series pilot data
- Support of scikit-learn pipelines for the data processing, as well as the pilot evaluation components
The components have been built on top of TensorFlow, scikit-learn, pandas, NumPy, and SciPy, while the library focuses
on state-of-the-art techniques for Time Series tasks, in both the data engineering and Neural Network models.
With the term "end-to-end AI pipelines", we mean actually, the development of a series of steps which has the following order: a) data ingestion (loading), b) data engineering, c) data preparation, d) AI model training (and fine-tuning), and, e) AI model interpretation and evaluation of its robustness and trustworthiness on pilot data, meaning data that emulates real-world scenarios.
The CrossAI library comprises the following components:
It provides processing tasks and functions that can be utilized both within and outside of pipelines for signal transformations and feature engineering. It also includes domain-based functionalities for audio, motion, etc.
It offers a collection of State-of-the-Art Neural Network architectures, built on top of
TensorFlow. These model topologies are designed to analyze Time Series data within both the 1D and 2D domains.
It delivers Explainable AI (XAI) mechanisms for interpreting the AI models' exports to actual exports of interest, as well as functionalities and algorithms for measuring the models' Trustworthiness and Robustness.
The implementation has been applied to the following study:
Tzamalis, Pantelis, Andreas Bardoutsos, Dimitris Markantonatos, Christoforos Raptopoulos, Sotiris Nikoletseas, Xenophon Aggelides, and Nikos Papadopoulos. "End-to-end Gesture Recognition Framework for the Identification of Allergic Rhinitis Symptoms." In 2022 18th International Conference on Distributed Computing in Sensor Systems (DCOSS), pp. 25-34. IEEE, 2022.
Next steps: To add Generative AI State-of-the-Art Neural Network Topologies.
If you use CrossAI in a scientific publication, we would appreciate citations, using the following Bibtex entry:
@misc{,
author = "{Tzamalis, Pantelis and Kapetanidis Panagiotis, and Kalioras, Fotios and Kontogiannis, George and
Stamatelos, John and Bardoutos, Andreas and Nikoletseas, Sotiris}",
title = "CrossAI: The Open Source library for End-to-End Time Series Analysis",
url = "https://github.com/AIoT-Group-UoP/crossai",
}
The following individuals are currently core contributors to the development and maintenance of the library:
The following people have been active in the contribution of previous versions of the library but no longer are engaging in its development:
The main design and idea have been created by Pantelis Tzamalis, as well as Andreas Bardoutsos in the early versions of the library, and occurred from needs during the development of AI projects with regard to Time Series tasks (classification, regression, clustering, etc.).
The main part of the library is developed and maintained by the Artificial Intelligence of Things Group (AIoT Group) in the Internet of Things Laboratory (IoT Lab) at the Computer Engineering and Informatics Department, University of Patras, Greece.
The library has been used, tested, and validated on various Research Projects funded by the EU and the Greek state, as well as projects funded by the Industry:
- Personal Allergy Tracer - National Project
- SynAir-G - EU Horizon
- Voice-Based Diagnostics - Private with Pfizer CDI
- Voice-Based Diagnostics 2 - Private with Pfizer CDI