Many techniques for relational learning construct features by aggregation, propositionalization, or other methods.
Some approaches to feature construction can cause biases in feature selection. For example, 
Jensen & Neville (2002) show that flattening some types of relational data can overestimate the amount of statistical evidence available to assess the utility of a feature. Other recent work (Jensen, Neville, & Hay 2003) shows that common aggregation functions can cause algorithms to mistake a type of structural regularity (degree disparity) for correlation among variables on different types of entities.