<p>The adoption of cloud-native microservices significantly improves system flexibility but introduces critical challenges in ensuring Quality of Service (QoS) and optimizing resource utilization in complex environments. A key obstacle lies in the scarcity of labeled performance data, as collecting and annotating microservice performance often requires extra profiling tools, thus weaking the effectiveness of the whole system. Moreover, resource allocation in such environments is inherently an NP-hard multi-objective optimization problem, but traditional heuristic methods easily fall into local optimum and poor generalization, failing to allocate the optimized resources adaptively. Thus, we focus on cloud-native resource allocation with low performance data, and then propose a Semi-supervised Meta-learning Prediction and Hyper-heuristic Allocation Model (SMPHAM). Firstly, SMPHAM adopts Generative Adversarial Networks (GANs) to augment limited labeled data, which will then be trained with Prototypical Network, a meta-learning framework, to improve the microservices performance prediction accuracy under low-data scenarios. Then, we allocate the resources according to prediction results by proposing a hyper-heuristic algorithm which adopts Ant Colony-based higher-level strategy to orchestrate three lower-level optimizers (Whale Optimization, Honey Badger Algorithm, and Northern Goshawk Optimization) to achieve a balance between QoS-requirement and resources limitation. Experiments validate that SMPHAM outperforms baselines in scarce performance data scenarios, and achieves superior prediction precision and resource efficiency.</p>