Class imbalance is recognized as one of the top 10 challenges in data mining and remains a major factor degrading the performance of downstream models. While many data-driven approaches attempt to mitigate this issue by generating synthetic samples, they often introduce unrealistic or mislabeled instances near decision boundaries. Counterfactual explanations provide a promising alternative by naturally generating near-boundary, label-flipping instances that help refine the decision boundaries. In this work, we propose NICE, a novel native instance-based counterfactual generation framework that constructs high-quality counterfactuals by adaptively combining existing instances with real feature values, rather than relying on interpolation. NICE identifies class-specific causal features and performs propensity score-based counterfactual matching for each causal feature to generate plausible minority-class instances, followed by filtering to remove low-quality samples. Experiments on 6 popular datasets demonstrate that NICE outperforms both widely used and counterfactual-based methods in addressing class imbalance.