Automatic Feature Engineering
In order to make machine learning techniques accessible to domain experts - the democratization of AI - one critical aspect is automatic, interpretable, and fast feature engineering. Representation learning methods mostly leverage neural networks to learn an embedding that results in a high accuracy for the corresponding machine learning task. However, such embeddings are often hard to interpret for users, which is in many cases essential in industry. Therefore, we focus on finding a feature representation that achieves both high accuracy and high interpretability. Additionally, the user can constrain the feature representation by construction time, size, fairness, and stability against errors. To solve this problem of finding the right feature representation - the needle in the haystack - we extend state-of-the-art feature selection methods by applying advanced sampling methods, such as instance selection, active learning, and meta-learning.