Topic model for online communities’ interests prediction
Amir Uteuov · Procedia Computer Science · 2019
This paper introduces the investigation of users interests by topic modeling. This work proposes a methodology for end-to-end unsupervised users topics investigation by textual data. Using additive regularization of topic models with large topic counts, tokens frequency filtering, lemmatization and communities information, one found interpretable interests. This paper shows differences of extracted topics from users posts and communities ones, from public user posts and hidden information from subscriptions. These topics differ significantly. This paper proposes a topics evaluation method by testing model prediction on the semi-automated labeled dataset. To get a more stable topic model one suggests using an averaged topic distribution based on users subscriptions. For cases with a large number of topics, one proposes an auto labeling approach solving as a classification problem with a constant number of classes.