We propose a method to predict users' interests by exploiting their various actions in social media. Actions performed by users in social media such as Twitter and Facebook have a fundamental property: user action involves multiple entities - e.g. sharing URLs with friends, bookmarking and tagging web pages, clicking a favorite button on a friend's post etc. Consequently, it is appropriate to represent each user's action at some point in time as a higher-order relation. We propose ActionGraph, a novel graph representation to model users' higher-order actions. Each action performed by a user at some time point is represented by an action node. ActionGraph is a bipartite graph whose edges connect an action node to its involving entities, referred to as object nodes. Using real-world social media data, we empirically justify the proposed graph structure. We show that the prediction accuracy can be improved by adequately aggregating various actions. Moreover, our experimental results show that the proposed ActionGraph outperforms several baselines, including standard tensor analysis PARAFAC, a previously proposed state-of-the-art LDA-based method and other graph-based variants, in a user interest prediction task. Although a lot of research have been conducted to capture similarity between users or between users and resources by using graph, our paper indicates that an important factor for the prediction performance of the graph mining algorithm is the choice of the graph itself. In particular, our result indicates that in order to predict users activities, adding more specific information about users activities such as types of activities makes the graph mining algorithm more effective.