In supervised machine learning, one of the major obstacles is data labeling, especially in fields such as academic record analysis, which require accurate annotation for structured data, and where the scarcity of labeled samples can impede model training. We present SALIENT (Self-supervised, Active, Learning, Integrated, Embedding, Neural, Transformation), a groundbreaking framework that significantly reduces the dependency on labeled data. This is achieved by incorporating self-supervised representations, active learning, and neural adaptation into the process. The approach utilizes a six-stage pipeline, comprising autoencoder-driven feature representation, centroid transformation, freezing pre-trained Tri-Training models, initializing an adaptation layer, followed by bootstrapping and pool-based active learning with an enhanced class balancing approach, tested on a confidential dataset containing 36,930 unlabeled academic records. Across five tests, SALIENT performance was remarkable; in Test 5, the model achieved 92% in accuracy with only 309 oracle queries. This outshines methods such as Tri-Training and Causal Inference, despite having higher accuracy, the human oracle labels were reduced by 77.5%. The framework is facilitated by its ability to rapidly cluster data and achieve quick convergence, greatly reducing the need for manual data labeling.
Badine et al. (Mon,) studied this question.