Published: 2020 Juni
Automated Machine Learning (AutoML) is the challenge of finding machine learning models with high predictive performance without the need for specialized data scientists. Existing approaches optimize a pipeline of pre-processing, feature engineering, model selection and hyperparameter optimization, and assume that the user is fully aware of the choice of the underlying loss function (such as precision, recall or F1-measure). However, end-users are often unaware of the actual implications of choosing a loss function, as the resulting models often significantly vary in their predictions. In this work, we propose a framework to personalise AutoML for individual end-users by learning a designated ranking model from pairwise user preferences and using the latter as the loss function for with state-of-the-art AutoML systems. Given a set of possible loss functions, we generate candidate models by repeatedly running AutoML with combinations of the former and have the user choose between pairs of resulting models. We use RankNet to learn a per- sonalised ranking function for the end-user, which is used as loss function for final run of standard AutoML system. To evaluate our proposed framework we define three preferences a user could pursue and show that a ranking model is able to learn these preferences from pairwise comparisons. Furthermore, by changing the loss function of AutoML we show that a personalized preference is able to improve machine learning pipelines. We evaluated the ability of learning a personalized preference and the entire framework on several OpenML multi-class classification datasets.