The detection of mental disorders through social media has received significant attention. With the growing prevalence of Autism Spectrum Disorder (ASD) and the inherent difficulties in diagnosing adults, researchers have attempted to identify undiagnosed adults. Previous studies have primarily concentrated on analyzing ASD characteristics rather than directly detecting ASD. The current study aims to propose a novel framework to assist in identifying the “lost generation” of ASD adults using their social media posts. Combining traditional and deep learning methods makes it possible to model complex aspects of ASD diagnostic characteristics, which have been relatively overlooked in previous studies. To accomplish this, specific formalizations for users’ patterns of interest as a main ASD diagnostic characteristic are proposed first. The latent linguistic and semantic features of ASD users’ postings are then modeled using deep and transformer-based language models. Finally, all these different aspects are considered together to train a detection model by employing the multi-view learning approach. The experiments show that the feature of idiosyncratic interests has more discriminative power than limited and repetitive interests. The results also indicate that the early fusion of interest-related features along with deep linguistic features outperforms the other examined feature combinations. Additionally, the proposed ‘’ fusion model demonstrates improved performance in capturing patterns of interests, compared to baselines. These findings suggest the potential application of the proposed framework towards indirectly identifying ASD users on social media, as evidenced by achieving precision and recall rates of 85% and 82% respectively on the used sampled dataset.

Keywords