 In brief, we use machine learning to study the characteristics of a small group of depressed people on Chinese social media called Weibo, which is actually a Chinese version of Twitter. And by studying their behavioral attributes, we developed classifier to estimate the risk of depression among the general population on social media. Firstly, we find that people with depression have different linguistic styles compared to people without depression. They are more frequent to express their mood in active words such as lonely, sad, scary, things like that. And secondly, we found out that they also have different diureal activities. Actually, the message they post on the social media takes late at night. That's between 11 p.m. to 3 a.m., actually. And they also have lower number of friends and followers on their social media, which means they actually have like ego network. And also, finally, 60% of the predict depressive individuals are women. Because we use the classifier to predict depressed people before they are actually on set, so we can offer this kind of information to the non-governmental organizations such as Lifeline to help them better, give them support. And also, because we like positively recognizing the people with depression, so we can help those with lower intention help seeking as well.