Existing graph pattern mining paradigms focus on repeated substructures or event orders and therefore overlook collaborative behavior, where correlated activities occur within short time intervals on structurally cohesive neighborhoods. We formalize this as Collaborative Pattern Mining (CPM), defining collaboration by joint temporal and structural proximity. We propose the support measure Collaborative Minimum Image Support (CMIS), defined as the minimum image count across participating labels, which is anti-monotone under the joint constraints and enables effective pruning. Building on this, we design piCPM, combining CMIS-based pruning with a partition-based index (PBI) constructed via diameter-constrained label propagation (DLPA) and augmented with precomputed multi-hop distance labels for fast graph queries. Across four real-world and synthetic datasets, piCPM uncovers meaningful dependencies among topology, attributes, and time, and it consistently outperforms strong baselines in accuracy and scalability, achieving up to three orders of magnitude speedups. These results establish CPM as a practical and scalable approach.