A method for deciding which subject domain a social-media user is genuinely interested in, by grounding the text they produce in a curated domain ontology rather than in surface word statistics. Each user's tweets and profile are mapped onto the classes and instances of a domain ontology (here a BBC Politics ontology extended with Australian parties, politicians and electorates), enriched through WordNet synsets and linked-data lookups via a commercial entity-annotation service, and the count of DISTINCT domain entities the user mentions per time period becomes the dominant prediction feature alongside platform metadata such as tweet, follower, friend, favourite, reply and retweet counts and verified status. A supervised classifier is then trained on hand-labelled users to label each account on-topic or off-topic for the domain. The point of the construction is that ontology-derived entity counts carry domain semantics that frequency-based topic models do not: on the same corpus a one-topic LDA or LSA returns only frequent surface words, whereas the ontology mapping recovers the high-level topic itself, and a decision tree over these features separates politically-interested users from the rest far better than a support vector machine over the same features.
Chapter 5 of a book on social big data analytics by Abu-Salih, Wongthongtham, Zhu, Chan and Rudra. Social Big Data (SBD) is the social-media content and metadata organisations accumulate (tweets, blo…