Complex Aggregates over Clusters of Elements

Abstract

Abstract. Complex aggregates have been proposed as a way to bridge the gap between approaches that handle sets by imposing conditions on specific elements, and approaches that handle them by imposing conditions on aggregated values. A complex aggregate summarises a subset of the elements in a set, where this subset is defined by conditions on the attribute values. In this paper, we present a new type of complex aggregate, where this subset is defined to be a cluster of the set. This is useful if subsets that are relevant for the task at hand are difficult to describe in terms of attribute conditions. This work is motivated from the analysis of flow cytometry data, where the sets are cells, and the subsets are cell populations. We describe two approaches to aggregate over clusters on an abstract level, and validate one of them empirically, motivating future research in this direction.

Counts-of-counts similarity for prediction and search in relational data

“…These frameworks typically extend classic machine learning models such as decision trees to operate on features extracted from relational neighborhoods. Most similar in spirit to our counts-of-counts features are perhaps complex aggregate features as considered in (Assche et al, 2006;Vens et al, 2014). These approaches do not define similarity measures on graph entities, and therefore can not be used for information retrieval tasks.…”