In assistive technology, misclassification can degrade performance by aiding the wrong activity. In human activity recognition classes are often significantly imbalanced and have few labels. For example, while the final stage of shooting is a continuous activity, often the only labeled samples are trigger pulls which can account for < 1% of the dataset. Label spreading can mitigate this but requires optimized features and balanced class distributions. Conventional metrics assume constant ground truth and work poorly with label spreading. This results in significant overrepresentation or loss of the minority class. In our proposed hybrid loop method, initial features are selected using correlations based on distance from the next shot, and continuous activities are used as a constraint. An iterative approach is taken in each step; Boruta is used to find important features based on current label distributions prior to label spreading. Transition Ratio is proposed as a metric based on the intuition that there should be a 1 to 1 ratio between labeled samples and class transitions. Motion and eye tracking data was collected from 10 subjects performing target acquisition and shooting tasks in virtual reality. 5-Fold cross-validation was used to compare baseline to hybrid loop. The hybrid loop decreased the size of the shooting class by 11.69 ± 0.89% without increasing misclassification of Trigger Pulls. Transition Ratio was 0.96 ± 0.06; nearly double those of the baseline models. This implies that the hybrid loop prevented over-representation of the minority class and focused its labels temporally around the endpoint labels. Better insight into the users’ current activities should improve contextual assistance.