12-02, 11:30–12:00 (UTC), Talk Track I
Understanding dependencies between features is crucial in the process of developing and interpreting black-box ML models. Mistreating or neglecting this aspect can lead to incorrect conclusions and, consequentially, sub-optimal or wrong decisions leading to financial losses or other undesired outcomes. Many common approaches to explain ML models – as simple as feature importance or more advanced methods such as SHAP – can yield misleading results if mutual feature dependencies are not taken into account.
In this talk we present FACET 2.0 - a new approach for global feature explanations using a new technique called SHAP vector projection, open-sourced at: https://github.com/BCG-Gamma/facet/.
Common black-box ML models do not provide explicit analysis of feature dependencies. This aspect of model interpretation is often neglected, which can lead to incorrect conclusions about the true contributions of features to individual predictions, or to the model as a whole.
FACET is an open-source ML library, developed by BGC Gamma with v2.0 about to be released with major enhancements. FACET uses a novel algorithm for global explanations of feature dependencies, addressing an important gap in existing approaches for black-box explanation methods. It introduces a measure of „redundancy” (how much information present in a feature is repeated in the other ones) and „synergy” (how much a given feature gains in predictive power when combined with other features).
Moreover, FACET is based on sklearndf, a library enhancing scikit-learn for full support of pandas DataFrames and keeping track of feature names across even complex ML pipelines.
Previous knowledge expected
I'm an AI Software Engineer working at BCG Gamma, passionate about ML, data science, functional programming and software excellence. In my free time I enjoy swimming and hiking.
Computer scientist specialising in Artificial Intelligence, Machine Learning & Software Engineering. Co-founder of BCG GAMMA and member of GAMMA's leadership.