Session 6: Cascaded Classification and Regression
Johannes Gutenberg-Universität Mainz
ABSTRACT CONTENT / DETAILS:
I will present new approaches to predicting the activity/non-activity of a chemical and the degree its activity in an integrated fashion, without having to define a combined objective function for classification and regression.
For the task of classification, the approach first learns a regression model for predicting the degree of activity and then uses the predicted degree of activity as a feature for the classification problem. We call this approach Regression Classification Cascade (RCC).
If the task is to improve the quality of a regression model instead, the opposite approach, called Classification Regression Cascade (CRC), is possible. In this approach, a classification model is first learned for discriminating active from inactive compounds.
The predicted classification is then used as an additional feature for improving the regression. In this setting, the inactive compounds serve as extra information besides the compounds along with their known activities. In comprehensive experiments, we show that both approaches can be beneficial across a number of different endpoints and data sets.
When used together with k-Nearest Neighbor approaches, it can also be useful in a read-across like context. Finally, we will discuss the potential of both approaches for elucidating mechanistic hypotheses.