Causal discovery in causal learning aims to derive causal graphs from observational data. However, collecting natural data is challenging and costly, so prior research has predominantly relied on synthetic datasets for validation. These synthetic or semi-real datasets, controlled artificially, may not fully reflect an algorithm’s performance in real-world scenarios. Therefore, we proposed a method for evaluating causal discovery in the absence of a ground truth causal graph. First, we divided the data into training and test sets, then performed causal discovery on the training set to obtain causal graphs. We subsequently conducted Markov blanket tests on the causal graphs using the test set and determined the causal direction of each edge using multiple methods. Finally, we integrated the results of these methods through weighted voting to achieve the final accuracy. Experiments on both synthetic and real datasets demonstrated that our proposed method can reflect true accuracy to some extent.