Memetic search for overlapping topics based on a local evaluation of link communities

摘要

In spite of recent advances in field delineation methods, bibliometriciansstill don't know the extent to which their topic detection algorithmsreconstruct `ground truths', i.e. thematic structures in the scientificliterature. In this paper, we demonstrate a new approach to the delineation ofthematic structures that attempts to match the algorithm to theoreticallyderived and empirically observed properties all thematic structures have incommon. We cluster citation links rather than publication nodes, usepredominantly local information and search for communities of links startingfrom seed subgraphs in order to allow for pervasive overlaps of topics. Weevaluate sets of links with a new cost function and assume that local minima inthe cost landscape correspond to link communities. Because this cost landscapehas many local minima we define a valid community as the community with thelowest minimum within a certain range. Since finding all valid communities isimpossible for large networks, we designed a memetic algorithm that combinesprobabilistic evolutionary strategies with deterministic local searches. Weapply our approach to a network of about 15,000 Astronomy & Astrophysics paperspublished 2010 and their cited sources, and to a network of about 100,000Astronomy & Astrophysics papers (published 2003--2010) which are linked throughdirect citations.