Social media posts often reflect users’ emotions and thoughts, sometimes containing subtle yet critical signals of suicidal tendencies. Traditional feature engineering-based methods often miss concealed distress. Although time-aware sequence-based methods are effective at learning and aggregating latent features from a user’s historical emotional spectrum, they frequently struggle to explicitly capture variations in the degree of suicidality and the presence of suicide-related associations across a user’s sequence of posts. A more effective approach would involve a deeper integration of both latent emotional historic context and textual features tailored to the specified user. In this paper, we propose an innovative self-adaptive Feature Learning framework (named FeaLearner), which transforms the suicide risk detection task into self-adaptive feature learning and selection procedures: firstly, it leverages a BERT-BiLSTM model to track users’ psychological-emotional evolution, enhanced by a temporal attention mechanism to highlight emotional historic context features. Furthermore, a multi-view adaptive feature selection network dynamically weights suicide-related textual features by integrating diverse sub-network perspectives, improving optimal text feature selection. Finally, the framework integrates emotional historic and textual context feature representations, and formulates an ordinal regression problem for suicide risk-level prediction. Experimental results demonstrate that FeaLearner yields better performance compared to various competitive baselines.