First, service compositions are specific types of objects that can vary in size (measured, for instance, by the number of services contained) and structure (e.g., how the services are connected). These objects can not easily be mapped to a feature representation (a vector of constant length) as typically requested by standard machine learning methods. To address this challenge, we develop the "learning-to-aggregate" framework, which is a novel setting of supervised learning. In this setting, we proceed from the assumption that the overall evaluation of a composition is an aggregation of the evaluation of its constituents. Thus, given a set of data in the form of compositions together with their evaluations, an essential part of the learning problem is to figure out the underlying aggregation function. Moreover, since the evaluations of local constituents are normally not observed directly, the problem also involves the "disaggregation" of the overall evaluation, i.e., hypothesizing on the evaluation of individual constituents. For example, in the case of reputation, an OTF user will provide a rating of the whole composition (as a single product), but not to all individual components. The overall rating then depends on how the user evaluates the individual components, and how she aggregates these evaluations into an overall rating (depending on the structure of the composition). The learning-to-aggregate approach seeks to answer both questions simultaneously .