Schema Independence of Learning Algorithms

Learning novel concepts and relations from structured data sets, such as relational databases, is an important problem with many applications in data management and machine learning. It is well established that the same data set may be represented under different schemas due to various reasons, such as efficiency, data quality, and usability. Further, the schema of a database may evolve over time. In this paper, we argue that relational learning algorithms should be schema independent, i.e. they should return basically the same results across various schemas of the same data set. Schema independent relational learning algorithms require less manual tuning and are easier to use over real-world data sets. We formally define and explore this property for different types of relational learning algorithms. We analyze the schema independence of some popular relational learning algorithms both theoretically and empirically. Our results indicate that these algorithms are not generally schema independent and will deliver different results and accuracies or require different amount of training data over different schemas for the same data set.