In many real-world scenarios, acquiring ground truth labels is expensive or impractical, making fully supervised learning infeasible. Weak supervision (WS) offers a practical alternative by employing user-defined heuristics or rules to approximate training labels in the absence of ground truth. In data programming, a widely used WS approach, these heuristics are implemented as labeling functions (LFs), which encode domain knowledge to assign noisy labels programmatically. However, the existing approach fails to fully exploit the domain knowledge encoded into LFs, especially when the LFs’ coverage is low, as the common data programming pipeline neglects to utilize data features during the generative process. This paper proposes a new method, reinforced labeling (RFL), which expands LF coverage by propagating their outputs to similar uncovered samples based on feature-space similarity. This improves label density without additional manual effort, enabling more effective training of end classifiers. Experiments on several domains (classification of YouTube comments, wine quality, and weather prediction) result in considerable gains over the data programming baseline, leading up to +21 points in accuracy and +61 points in F1 scores.