Overview

The Explabox aims to provide insights into your data and model behavior, by transforming ingestibles into digestibles through four types of analyses. The Explabox is split into three layers:

Ingestibles

Ingestibles encompass your model and data. The Ingestible class provides a unified interface for accessing your model (e.g. its meta-information and prediction function) and your data (e.g. the instances, the various splits, and corresponding ground-truth labels). Ingestibles contain meta-information of the data and model (to determine relevant functions for analyses and for auditability) and allow for optimized inferencing through batching and lazy loading.

Importing ingestibles

The model can be imported via the import_model() function, while the data is imported via the import_data() function.

Supported models are:

• Scikit-learn

• ONNX (supports TensorFlow, PyTorch, HuggingFace, Scikit-Learn, …)

• Python callables

Supported data are:

• NumPy arrays

• Pandas DataFrames

• HuggingFace datasets

• Files (online or offline):

  • Tabular dataset files (.csv, .tsv, .json, .txt, .xls, .xlsx, …)

  • Folders or archived files (.zip, .rar, .gzip, …) containing multiple dataset files

  • Pickled Python objects (.pkl)

  • HDF5 files

Analyses: turning ingestibles into digestibles

Once imported, these ingestibles can be made more informative by turning them into digestibles. Each analysis provides functions to enhance the transparency of the model and/or data:

Analysis	Class	Description	Requires
explore	Explorer	The Explorer explores your data by providing descriptive statistics.	data
examine	Examiner	The Examiner calculates quantitative metrics on how the model performs.	data, model
explain	Explainer	The Explainer creates explanations corresponding to a model and dataset (with ground-truth labels).	data, model
expose	Exposer	The Exposer exposes your model and/or data, by performing sensitivity tests.	data, model

The Explabox class is a unified interface for all of these classes, where an instance of each of these classes is constructed. These can be accessed via .explore, .examine, .explain and .expose, respectively.

For example, using the Examiner one can obtain (for a classification task) all instances wrongly classified by the model, which returns the WronglyClassified digestible:

>>> from explabox.examine import Examiner
>>> Examiner(ingestibles=ingestible).wrongly_classified()

Digestibles

Digestibles are the return type after performing an analysis. The information contained in them can be accessed in various ways (i.e. interactively or statically), depending on stakeholders needs. Example methods of access include:

• Python object (i.e. the raw property values)

• Object descriptions:

  • JSON

  • YAML

  • Python dictionary

• User interfaces (using plots rendered with plotly):

  • HTML

  • Jupyter Notebook

When using an interactive Python shell, the Explabox will default to the Jupyter Notebook interface. In a non-interactive Python shell the Python object is the basis for the digestible. For online applications (e.g. streamlit) the HTML interface can be used.