Controlling collusion in government bidding is a prerequisite for ensuring social justice and the smooth operation of projects. Based on the prospect theory, this article establishes a four-party evolutionary game model for tenderers, enterprises with higher willingness to collude, enterprises with lower willingness to collude, and supervising enterprises. The study uses replication dynamics to analyze the stability of strategy selection after the evolutionary game. The results show that higher project base returns increase the probability of collusion, while lower market competition, higher risk aversion, and stronger collusive regulation all reduce the probability of collusion. When regulators adopt a strong regulatory strategy, the remaining project participants tend to choose a noncollusive strategy.

The classification of collusion behaviors of government-invested project tenderers is one of the important methods to describe the characteristics and laws of collusion behaviors and strengthen the governance of collusion. Firstly, the variables that affect the type of collusion behavior are selected and cluster analysis is carried out on the cases of collusion in government investment project bidding. Then use the social network to mine the types and characteristics of the collusion behavior of the tenderee. Finally, a BP neural network automatic identification model is established to quickly discriminate the types of collusion, which can effectively overcome the subjectivity of traditional methods. As a result, three typical types of collusion of tenderees can be obtained: intervention type, opportunity type, and cooperation type. The study found that the three types of collusion behavior have their own characteristics and laws, and the relationship between the variables that affect the type of collusion behavior is complex and affects each other.