Data-Mining of Social Media Users with Embedding Techniques and Neural Network

In this study of ten thousand representative users from the popular Chinese social media platform bilibili.com, we find that usernames—ordinarily considered mere indexing elements rather than useful modeling features — alongside personalized signatures and other natural language data, should and must be integrated into neural network modeling via embedding techniques, leveraging pretrained large language models such as BERT. This work challenges conventional wisdom and emphasizes the importance of harnessing all potentially informative user-generated content, even those previously deemed extraneous. The successful application of embedding methods to these unconventional data sources underscores their versatility in extracting meaning from diverse text types, extending the scope of text-based feature engineering. This discovery holds promise for uncovering subtle user traits, enriching user profiles, and guiding practical applications such as user engagement optimization, content personalization, and targeted marketing efforts. Overall, our study advances the understanding of the value of usernames and signatures as valuable modeling inputs, contributing to the refinement of social media analytics and user profiling.