Type Extension Trees for feature construction and learning in relational domains

Abstract

Please cite this article in press as: M. Jaeger et al., Type extension trees for feature construction and learning in relational domains, Artificial Intelligence (2013), http://dx.doi.org/10. 1016/j.artint.2013.08.002 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Type Extension Trees for Feature Construction and Learning in Relational DomainsManfred Jaeger AbstractType Extension Trees are a powerful representation language for "count-ofcount" features characterizing the combinatorial structure of neighborhoods of entities in relational domains. In this paper we present a learning algorithm for Type Extension Trees (TET) that discovers informative count-of-count features in the supervised learning setting. Experiments on bibliographic data show that TET-learning is able to discover the count-of-count feature underlying the definition of the h-index, and the inverse document frequency feature commonly used in information retrieval. We also introduce a metric on TET feature values. This metric is defined as a recursive application of the Wasserstein-Kantorovich metric. Experiments with a k-NN classifier show that exploiting the recursive count-of-count statistics encoded in TET values improves classification accuracy over alternative methods based on

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

“…Our approach to defining similarity between graph objects is feature based: we use a highly expressive and flexible framework for defining features for graph entities, and then compare entities by comparing their feature values, based on suitable feature metrics. For feature definition, we use the type extension tree (TET) framework, which we introduced in (Jaeger et al, 2013).…”

“…Our approach is distinguished by the fact that it takes complex quantitative aspects of relational neighborhoods into account. Specifically, we base our distance metrics on counts-of-counts features we introduced in (Jaeger et al, 2013), that represent detailed quantitative information about relational neighborhoods. To illustrate the basic idea of counts-of-counts features, consider the case of a bibliographic database containing authors and papers, an author of relation between authors and papers, and a cites relation between papers.…”

“…The numbers underneath the graphs for Graham and Curtis are the actual distances to Bogart according to our genre-based and business-based metrics. We use our earlier type extension tree framework (Jaeger et al, 2013) for specifying suitable counts-of-counts features. This paper makes the following new contributions:…”

“…A simple precursor of the metric introduced in this paper was already described in (Jaeger et al, 2013). That metric was defined directly on the raw counts-of-counts feature values, and did not permit the kind of problem-specific adaptation we now obtain through the logistic evaluation function.…”

Complex Aggregates over Clusters of Elements

“…Other types of complex aggregates that have been proposed include count-ofcount features, which are a kind of nested aggregate, and have been introduced by [9,11] in the type extension tree representation language. An example of such a feature is the complete specification of how many atoms that are present in the molecule to be classified are bound to how many atoms with an aromatic bond type.…”