This study explores collusion detection in procurement auctions at a private very large industrial park in Taiwan using unsupervised methods. By analyzing big data and applying graph theory, contractor relationships are modeled as dynamic networks, with nodes representing contractors and edges indicating shared workforce or resources. A community detection algorithm, Girvan-Newman, identify cooperative and competitive patterns. Screening variables like coefficient of variation, spread, difference percentage, relative distance, skewness, and Kolmogorov-Smirnov test are used to detect irregularities. While logistic regression shows stable performance with 70% accuracy, advanced models like XGBoost and CatBoost achieve higher accuracy but exhibit overfitting. The dynamic contractor analysis aids procurement managers in identifying collusive behavior and recommending competitive contractors. It is able to enhance fairness and efficiency in the procurement auctions.