Ideology detection is the problem targeting at telling people’s ideal point. To be specific, the goal is to predict their political tendencies, measured by which party they support. It has long been a challenging yet important problem. Making precise estimation on the lawmakers’ ideal point will help us foresee the results of the voting of the bills, and precisely estimating the ideal point of ordinary citizens would help politicians / news agencies make wise decisions.Researchers have been working on ideology detection of the law-makers and congressman since decades ago, but analyzing ordinary citizens’ political tendencies was uneasy, requiring labor-intensive survey-study. Even if we consider the rise of social networks likeTwitter that attracts massive amount of users with diverse back-grounds from around the globe, making efficient and effective use of the social media data could still be a problem. The lack of labels and missing features introduce a lot of challenges on their own, the size of the graph could make things harder. All those aspects differ at a fundamental level from the commonly-used datasets the state-of-the-art heterogeneous network embedding models applied to. We crawled our own datasets from Twitter, and proposed TIMME, a multi-task multi-relational embedding model. First, we overcome the problems brought by label-sparsity by developing a multi-task training framework. Second, we altered the GCN layer to handle multiple relations while keeping its advantages such as high efficiency. Third, we also propose a unique design of handling the missing features by treating them as parameters to be learned. Experimental results showed that TIMME is overall better than the state-of-the-art models for ideology detection onTwitter.