Machine learning models are recognized as having high predictive power for data in which interaction effects are likely to be strong, but such models tend to be black boxes that are difficult to interpret and lack explanatory power in prediction. Therefore, an urgent issue is to develop methods to help interpretation by visualizing and measuring the interaction effects captured by machine learning models, and interpretable machine learning techniques that are expected to be used for this purpose have been developed. However, the two existing visualization methods, partial dependence and accumulated local effects, are inadequate because they fail to capture the main effects well, let alone the interaction effects, even for relatively simple models. In this paper, we propose alternative methods for understanding interaction effects as well as indicators for evaluating these methods. For this purpose, we define interaction effect terms and identify their various properties. Using the proposed indicators, we also characterize existing and alternative methods for simple cases. These analyses and examinations show what is lacking in specific techniques for visualizing and measuring interaction effects. We also recognize strongly that it is necessary to discuss interaction effects from basic principles.

Find the Q&A here: Q&A on' Machine Learning and its Opportunities'