Unsupervised User Identity Linkage from User-Generated Text in Sparse-Signal Settings

User identity linkage (UIL) aims to determine whether different online accounts, profiles, or aliases correspond to the same underlying individual. Existing methods often rely on profile, social, or temporal signals that may be sparse, inaccessible, or non-comparable across contexts. This motivates the study of UIL using textual content alone, framing it as an unsupervised authorship clustering problem. We represent texts through complementary channels, including character and word n-grams, dense semantic embeddings, named-entity signals, and a custom stylometric-linguistic feature set to capture interpretable markers of writing style. These representations are combined through fusion schemes and evaluated with centroid-based clustering and graph-based community detection. Experiments on multiple datasets show that performance depends strongly on the choice of representation, clustering method, and dataset characteristics. Overall, the results indicate that unsupervised text-based UIL is a promising alternative when profile, graph, and temporal signals are unavailable.