Positive and Unlabeled Learning Techniques

Positive and unlabeled (PU) learning addresses binary classification when only confirmed positive instances and a pool of unlabeled data are available. This paradigm combines elements of supervised and semi-supervised learning to reduce the cost and difficulty of obtaining negative labels. Typical approaches estimate the proportion of positives within the unlabeled set or identify a subset of reliable negatives, and then train a classifier on the labelled positives alongside these inferred negatives. Advances in the field have introduced generative frameworks, neighbourhood-based algorithms, tensor-network models and quantum-inspired strategies, each seeking improvements in accuracy, computational efficiency and robustness to distributional assumptions. Real-world applications span medical diagnosis, drug discovery, environmental risk mapping and web mining, where negative labels are scarce or prohibitively expensive. Research continues to refine loss functions, sampling schemes and evaluation metrics to address challenges such as class prior estimation bias, label noise and high dimensionality.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Positive and Unlabeled Learning Techniques publication trend

The graph below shows the total number of articles in positive and unlabeled learning techniques across all publications each year (not limited to Nature Index journals).

Technical terms

Positive and Unlabeled (PU) Learning: A classification setting using only labelled positive examples and unlabelled data containing both positive and negative instances.

Class Prior: The true proportion of positive examples within the unlabeled set, essential for unbiased classifier training.

Reliable Negative Example: An unlabeled instance inferred with high confidence to belong to the negative class, used to complement positive labels in training.

Permutation Testing: A non-parametric method that assesses model performance by comparing results on true labels with those obtained under random label permutations.

Entropy Measure: A criterion for quantifying uncertainty or impurity in data partitions, commonly employed in decision-tree algorithms.

References

1. Conditional generative positive and unlabeled learning. Expert Systems with Applications (2023).
2. Positive unlabeled learning with tensor networks. Neurocomputing (2023).
3. A Novel Classification Method: Neighborhood-Based Positive Unlabeled Learning Using Decision Tree (NPULUD). Entropy (2024).
4. Leveraging permutation testing to assess confidence in positive-unlabeled learning applied to high-dimensional biological datasets. BMC Bioinformatics (2024).
5. Positive Unlabeled Learning Selected Not At Random (PULSNAR): class proportion estimation without the selected completely at random assumption. PeerJ Computer Science (2024).
6. A Quantum-Inspired Direct Learning Strategy for Positive and Unlabeled Data. International Journal of Computational Intelligence Systems (2023).