Cyber-physical-social systems (CPSSs) represent an emerging and complex system architecture that integrates the physical world, computational technologies, and social behaviors, demonstrating significant potential across various domains such as recommendation systems, Industry 4.0, smart healthcare, and intelligent transportation. However, as CPSSs are still in their developmental stages, security concerns remain relatively low. In CPSSs, the network serves as both a producer and transmitter of data, making its security paramount. Existing intrusion detection methods for CPSSs often overlook the unique characteristics of these systems. To address this gap, we propose the collaborative hybrid recursive feature addition (CHRFA) method, which combines hybrid recursive feature addition (HRFA) and a collaborative learning algorithm. We introduce a novel method for calculating feature redundancy, called redundancy sensitivity (RDS), which HRFA employs to efficiently assess redundancy and rank feature importance. In the collaborative learning algorithm, individual agents first utilize HRFA to independently rank all features. Subsequently, a central server consolidates the ranking results from all agents to generate a final feature ranking, thus enhancing the efficiency of feature selection. The experimental results show that HRFA achieves 99.79% and 99.92% accuracy on the CIC-IDS2017 and ISCX-IDS2012 datasets, improving upon Recursive Feature Addition (RFA) by 2.89% and 1.4%, respectively. Moreover, the time cost of CHRFA in the multi-agent setting is lower than that of RFA.