The comprehensive evaluation indicators of complex systems have a large selection space, and there are complex correlations between the evaluation indicators. How to reasonably select the evaluation indicator, that is, to optimize the indicators set, is an important premise for carrying out the comprehensive evaluation, and is the key to enhance the credibility and vitality of the comprehensive evaluation results. This article proposes an indicator set optimization technique based on the idea of bipartite graphs. Firstly, it takes evaluation indicators and their series of monitoring data as inputs and constructs an evaluation indicator network by measuring the vector distance between evaluation indicators. Then, the evaluation indicator selection problem is modeled as a network bipartite graph of the selected indicator set and the abandoned indicator set. Finally, by improving the KL algorithm, the indicator selection problem can be efficiently solved. We conducted an experiment on optimizing the evaluation index set based on the global firepower index. The results show that the proposed method could be well applied to the optimization problem of comprehensive evaluation indicator sets, and has high solving efficiency.