Original title: Focused PU learning from imbalanced data
Elias Zavitsanos, Georgios Paliouras · September 2026
In many machine learning tasks, positive examples are rare, and the rest of the data is completely unlabeled. This unlabeled group mostly contains negative examples, but it also hides a few positive ones. Real-world tasks like finding fraud or disease-related genes face this issue. Current tools often struggle when data is heavily imbalanced, especially when positive examples look a lot like negative ones. To solve this, researchers built a new learning method. It trains a two-group classifier using a focused risk estimator that carefully evaluates both the positive and unlabeled examples together. In tests, this system achieves state-of-the-art performance across imbalanced datasetsDatasets where one category (like fraud) appears much less often than the other (like normal transactions). under two different data-labeling conditions. It also proves useful in the real world for detecting financial misstatements.
Imagine you are a detective trying to find a rare counterfeit coin among thousands of real coins. You have a few confirmed counterfeits, but the rest are unlabeled. The problem is that the counterfeits are made to look exactly like real coins, so your usual detection methods fail. In this paper, the rare counterfeits are the positive examples, the unlabeled pile is the unlabeled data, and the method uses a focused risk estimator to pay extra attention to the coins that look most like real ones but might be fake.
Many real-world problems, like fraud detection or disease gene identification, have very few labeled positive examples, and those positives often look like negatives. Existing machine learning methods fail in these imbalanced settings. This new method improves detection of such rare, hard-to-spot positives, making it valuable for finance, healthcare, and other fields where missing a positive case is costly.
The abstract does not discuss limitations explicitly. However, the method is tested under specific labeling assumptions (SCAR and SAR) which may not hold in all real-world scenarios. The real-world application is limited to financial misstatement detection; performance in other domains is not shown. The focused risk estimator might need adjustment for different levels of imbalance or noise in the unlabeled data.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Focused PU learning from imbalanced data — Elias Zavitsanos, Georgios Paliouras — Data Mining and Knowledge Discovery — doi:10.1007/s10618-026-01264-1
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.