Knowledge graphs, as databases that represent knowledge in a graph structure, offer abundant resources for numerous applications. This paper is dedicated to the online aggregations tailored for knowledge graphs. Online aggregations refer to execute aggregate queries in real-time through factoid query processing and unbiased sampling, their efficiency and accuracy are often limited by the factoid queries. To overcome these hurdles, this paper proposes online aggregations incorporated with graph representations. We first design a graph representation model that fuses long-neighbor attention to provide better representations for the online aggregations on knowledge graph. Furthermore, a distribution feature elimination of representations for unbiased sampling and a two-stage sample pruning for irrelevant samples filtering are designed, thereby reducing the computation for the aggregate queries. By merging online aggregations with graph representations, an unbiased estimators are introduced, along with the confidence interval of the estimated result on the samples. Experimental results demonstrate that our model can efficiently return effective aggregate query results with guaranteed relative error, successfully tackling existing challenges for online aggregations on knowledge graphs.