To explain a complex machine learning model, we generate a local approximation (or surrogate model) that can be easily explained.
We built ExplainExplore: a new approach for analyzing and understanding classification models
using state of the art machine learning explanation techniques. The system offers three perspectives:
Individual instances are explained using feature contribution vectors extracted from surrogate models. This shows how much every feature contributed to the prediction.
To assert the correctness of these values, information on prediction certainty (P) and surrogate fit (R2) are shown.
In the context view, nearby data points are shown as well as the class probability of complex reference model (global) and surrogate (local). The expert can use this to assert whether the surrogate is locally faithful to the reference model, and explore the effect of instance perturbations.
The global overview combines the contribution vectors for all data in one Parallel Coordinate Plot (PCP). This helps the data scientist to identify whether the seleted instance was classified similar to other instances, and whether the model has 'strategies' (clusters in the contribution vectors).
ExplainExplore was built to support a wide variety of datasets and models. The system supports:
We created a limited version of ExplainExplore that runs completely in the browser (as opposed to
requiring a Python backend). Try it out below.
If you want to refer to our system, please cite our paper using the following BibTeX entry: