A1576
Title: Essence codings: A data-driven approach for interpreting black-box models through factorial effects
Authors: Chao-Hui Huang - Academia Sinica (Taiwan) [presenting]
Shao-Wei Cheng - National Tsing Hua University (Taiwan)
Abstract: Factorial designs are a common choice for conducting screening experiments. Traditionally, the codings for factorial effects are predetermined before the experimental phase begins. While this approach is simple and widely used, these predefined codings may fail to capture a substantial portion of the variability in the response, even when certain factors clearly have strong influence. A new data-driven coding system for factorial effects, referred to as essence codings, is proposed. The proposed approach constructs coding functions directly from the observed data, with the goal of capturing the dominant patterns of variation in the response. As a result, essence codings provide a more informative and interpretable representation of factor effects compared to traditional codings. Examples are presented to demonstrate the practical usefulness of the proposed method, including applications to the interpretation of black-box models. The results show that essence codings can effectively reveal meaningful structures in the data and facilitate model interpretation.
