Cultural scene detection using reverse Louvain optimization
This paper proposes a novel approach for discovering cultural scenes in social network data. "Cultural scenes" are aggregations of people with overlapping interests, whose loosely interacting activities form virtuous cycles that amplify cultural output (e.g., New York art scene, Silicon Valley startup scene, Seattle indie music scene). They are defined by time, place, topics, people and values. The positive socioeconomic impact of scenes draws public and private sector support to them. They could also become the focus for new digital services that fit their dynamics; but their loose, multidimensional nature makes it hard to determine their boundaries and community structure using standard social network analysis procedures. In this paper, we: (1) propose an ontology for representing cultural scenes, (2) map a dataset to the ontology, and (3) compare two methods for detecting scenes in the dataset. Method One takes a hard clustering approach. We derive three weighted, undirected graphs from three similarity analyses; linking people by topics, topics by people, and places by people. We partition each graph using Louvain optimization, overlap them, and let their inner join represent core scene elements. Method Two introduces a novel soft clustering approach. We create a "scene graph": a single, unweighted, directed graph including all three node classes (people, place, topic). We devise a new way to apply Louvain optimization to such a graph, and use filtering and fan-in/out analysis to identify the core. Both methods detect core clusters with precision, but the first method misses some peripherals. Method Two evinces better recall, advancing our knowledge about how to represent and analyse scenes. We use Louvain optimization recursively and in reverse to successfully find small clusters.