Iterative Two-Phase Active Learning for Skewed Data

The patent segments the large skewed data set into multiple smaller partitions or batches. Each partition is processed independently through the active learning loop, allowing parallel processing and reducing the overall time required for labeling while maintaining model training quality through systematic coverage of the entire data set.

The patent performs preliminary analysis to identify and prioritize regions of interest or high-value samples before full labeling begins. By pre-processing the data to flag potentially informative examples, the system reduces the time needed for manual annotation while ensuring that the most valuable data points are labeled first.

The patent implements an active learning system where the model iteratively selects its own training examples based on uncertainty sampling or other criteria. This self-service approach allows the system to automatically identify which samples require expert labeling, minimizing the time experts need to spend while maintaining high labeling accuracy for the most critical samples.

The patent applies partial labeling by having experts annotate only a subset of the most informative samples identified by the active learning algorithm, rather than labeling the entire data set. This partial action approach achieves sufficient labeling accuracy for model training while dramatically reducing the time investment required from subject matter experts.

The patent replaces random sampling with active learning algorithms that use feedback from the model's performance and uncertainty estimates to intelligently select the next batch of training examples. This feedback-driven approach maintains systematic simplicity while dramatically improving training efficiency by focusing computational resources on high-value samples that will most improve model performance.

The patent changes the sampling parameter from uniform random selection to probability-based selection weighted by sample informativeness, uncertainty, or predicted value. This parameter change transforms the sampling strategy from simple but inefficient random sampling to a more complex yet highly efficient selective sampling approach that prioritizes examples most likely to improve model performance.

The patent segments the annotation process into multiple iterative rounds, where each round processes a subset of data. This segmentation allows the system to achieve comprehensive model training by systematically covering different regions of the data distribution across multiple passes, while reducing resource consumption in each individual pass by working with smaller data subsets.

The patent applies partial annotation by selectively labeling only the most informative samples identified through active learning criteria, rather than annotating the entire data set. This partial action approach achieves sufficient training completeness for the target performance level while dramatically reducing computational resource consumption compared to full data set annotation.