Predictor Factory: Učení z relačních dat

Abstract

Propositionalization algorithms transform relational data into a single table with features, which can be used for classification or regression with conventional machine learning tools. However, the contemporary propositionalization algorithms were not designed to work on changing data and suffered from the production of many irrelevant and redundant features. We altered the propositionalization to work on temporal data and introduced meta-learning, which predicts, which features will be relevant and unique. To test Predictor Factory, our implementation of propositionalization, we have created a repository of relational datasets and implemented a scalable algorithm for relationship discovery between tables in the dataset. The implementations were open-sourced and applied to real-world banking, government, marketing, medicine, and telecommunication problems.

Propositionalization algorithms transform relational data into a single table with features, which can be used for classification or regression with conventional machine learning tools. However, the contemporary propositionalization algorithms were not designed to work on changing data and suffered from the production of many irrelevant and redundant features. We altered the propositionalization to work on temporal data and introduced meta-learning, which predicts, which features will be relevant and unique. To test Predictor Factory, our implementation of propositionalization, we have created a repository of relational datasets and implemented a scalable algorithm for relationship discovery between tables in the dataset. The implementations were open-sourced and applied to real-world banking, government, marketing, medicine, and telecommunication problems.