Year: 2026
Type: article
Abstract: Sociodemographic information provides critical context for interpreting digital behavior and its societal implications in many computational social science settings. However, the labels used to train and validate current inference methods frequently rely on researcher-defined proxies or opaque modeling assumptions. This can make demographic measurement difficult to audit, especially when inferred attributes are used to interpret population-level patterns. Here, we study sociodemographic inference on Reddit using over 850 000 public user self-declarations of age, gender, and partisan affiliation as an explicit source of supervision. We train models on sparse subreddit participation signals and compare simple probabilistic models with embedding-based baselines commonly used for Reddit analysis. Across coarse binary classification and aggregate quantification tasks, Multinomial Naive Bayes (NB) often performs better than more complex alternatives. Relative to the unsupervised state-of-the-art embedding method, NB improves ROC AUC by up to \(19\%\) . Under true supervision, it maintains quantification error below \(15\%\) for gender and year of birth. The models produce calibrated and interpretable outputs, thus supporting uncertainty estimation and subreddit-level feature importance analysis. These results position self-declarations as an explicit and auditable source of supervision for studying both user-level and population-level sociodemographic patterns on Reddit. By grounding labels in user-provided statements, the framework reduces reliance on researcher-selected proxies and supports transparent aggregate analyses of online communities.
Source: Scientific Reports
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W4407310258