Interpretability of neural networks predictions using Accumulated Local Effects as a model-agnostic method

• Published in: Computer Aided Chemical Engineering Vol. 51; pp. 1501 - 1506
• Main Authors: Danesh, Tina, Ouaret, Rachid, Floquet, Pascal, Negny, Stéphane
• Format: Book Chapter
• Language: English
• Published: 2022
• Subjects: Accumulated local effects; Interpretability; Model-agnostic method; Neural network; Model-agnostic method; Neural network; Accumulated local effects; Interpretability
• ISBN: 0323958796; 9780323958790
• ISSN: 1570-7946
• DOI: 10.1016/B978-0-323-95879-0.50251-4

Recently, machine learning methods such as neural networks have been applied in various applications thanks to their accuracy and flexibility. However, the main drawback of these methods is the lack of interpretability, which is the reason for being uncommon in chemical engineering applications. At the same time, the recent rise of interpretability research has led to some confusion in various communities. In order to deal with this issue for the CAPE community, we propose to discuss the notion of interpretability under the prism of neural network predictions using an example from chemical engineering. To do this, we first set out the framework for defining interpretability for machine learning methods. Then a post-hoc method for model evaluation, explicitly named ”model-agnostic methods”, will be presented. In this study, we try to enhance the interpretability of the neural network predictions and visualize the effect of features on the model’s output by a model-agnostic method named Accumulated Local Effects. As a case study, we work on predicting electrical power output and prioritizing the parameters of a combined cycle power plant. We could conclude that the most influential input parameter among Ambient Temperature (AT), Atmospheric Pressure (AP), Vacuum (V), and Relative Humidity (RH) is AT, and the most interaction is between AT and V.