Counting triangles in a graph is an important step to calculate the clustering coefficient and the transitivity ratio of the network. Under the background of big data

the sampling method is the main method used in triangle counting. However

two problems of spatiotemporal costs and computing errors are very difficult. An approximate triangle counting algorithm based on Markov chain in the graph stream is proposed. Treating the window as processing unit of graph stream

it uses Markov chain to implement dynamic adaptation of memory and to reduce error ratio while sampling. Compared with other algorithms

the proposed algorithm can drop the error rate by 2~4 times

and the time consumption has a great improvement.