Frequent Pattern Mining (FPM) on large graphs serves as a cornerstone in graph data mining, aiming to uncover patterns with support exceeding predefined thresholds. However, traditional FPM techniques often fall short of meeting users’ actual requirements, primarily due to their dependence on single-dimensional evaluation metric and inadequate consideration of subjective preferences. To tackle the issue, we introduce the Certified Pseudo-label enhanced Active Learning Framework (CPALF) for pattern interest evaluation, designed to accurately predict users’ subjective preferences on patterns. CPALF is characterized by three key features: (a) it leverages active learning to efficiently gather users’ preferences via limited human-computer interaction; (b) it integrates semi-supervised learning to produce high-confidence pseudo-labeled training samples from unlabeled data; and (c) it adopts a replay strategy to alleviate the forgetting problem in incremental learning. This approach significantly enhances prediction performance while reducing reliance on annotated data. Experimental results indicate that CPALF effectively captures users’ preferences, achieving up to 96% prediction accuracy with limited labeled data. The code is available at https://github.com/waistboo/CPALF .