The impact of attention mechanisms and deep learning on multi-class occupation prediction

Today, analyzing big data sources is becoming increasingly important for understanding individuals’ occupational, personal, and social characteristics. Social media platforms contribute to solving complex problems, such as occupational prediction, through text-based content that provides clues about users’ daily lives and professions. Such analyses are critical not only at the individual level but also for understanding sectoral and societal trends. In this study, several deep learning architectures were developed to predict both occupational groups (36 classes) and individual occupations (65 classes) using a dataset of approximately 500,000 Turkish X posts. In addition to LSTM, ImprovedLSTM, GRU, and RNN models, a total of 64 different models were created using architectures enriched with Attention mechanisms. For the 36-class occupational group prediction, the best model was LSTM, achieving an accuracy of 97.13%, while for the 65-class individual occupation prediction, the best model was the Attention-based GRU, which achieved an accuracy of 93.34%. The results of the study demonstrate that deep learning techniques offer more efficient and scalable solutions by reducing the need for preprocessing. It was also found that Attention mechanisms play an important role in maintaining model accuracy as the number of classes increases.