LeaDQ is a research method for choosing which examples from decentralized, unlabeled data streams should be sent for annotation. It uses multi-agent reinforcement learning to teach clients local query policies, with implicit global guidance intended to keep selections useful to the shared model. The authors report simulation results on image and text tasks, but the abstract gives no numeric effect size and does not establish performance in a live deployment.
Why querying unlabeled streams is difficult in federated learning
Many federated-learning studies assume that participating clients already have labeled examples. LeaDQ considers a different setting: data arrive over time at separate clients, but the incoming examples do not have ground-truth labels. Since annotation takes effort and resources, the system must choose which examples are worth labeling.
As an Amazon Associate I earn from qualifying purchases.
The challenge is not simply to find examples that appear informative to one client. A locally attractive example may contribute little to the shared model, while a sample useful to the global objective may not look especially valuable from one client’s perspective. Querying therefore has to balance local data and the needs of collaborative training.
How LeaDQ chooses examples
LeaDQ frames query selection as a collaborative decentralized decision problem. Its proposed mechanism is multi-agent reinforcement learning: client-level policies learn which stream examples to select for annotation, while implicit global information guides those choices toward samples that may help the shared model.
#1 Best Overall
The method alternates between local data querying and model training. In broad terms, clients use learned policies to select examples, those examples can be annotated and used for training, and the shared-model objective supplies guidance to the selection process. This describes the authors’ approach; it is not a guarantee that every selected sample will improve the global model or that the method will behave the same way under every distribution.
What the reported results establish
The AAAI paper’s abstract reports extensive simulations on image and text tasks and says LeaDQ improves performance over benchmark algorithms in various evaluated federated-learning scenarios. The evidence available in the abstract is qualitative: it does not provide a numeric improvement to quote here, and simulation findings alone do not demonstrate results in a production system.
Rank #2
The paper appeared in the Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, issue 19, pages 20752–20760, and was published April 11, 2025. It is by Yuchang Sun, Xinran Li, Tao Lin, and Jun Zhang. AAAI paper page · DOI record.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →How LeaDQ fits alongside other federated active-learning methods
Federated active learning covers approaches that select data for labeling in order to improve a shared model while limiting annotation. Related methods differ in their assumptions and selection mechanisms, so they should not be treated as interchangeable with LeaDQ.
| Method or context | Data and task setting | Selection approach and distinction |
|---|---|---|
| LeaDQ | Unlabeled data streams at federated clients; simulations include image and text tasks. | Multi-agent reinforcement learning with implicit global guidance; alternates querying and model training. AAAI paper page |
| LoGo | Federated active-learning setting; the cited work examines how inter-class diversity at local and global levels affects selector performance. | Combines global and local selectors in two selection steps. This is context for the local-versus-global selection problem, not LeaDQ’s method. CVPR Open Access paper |
| FALE | Federated active data selection for regression with non-IID clients; the proceedings abstract reports experiments on 11 benchmark datasets. | Uses leverage-score sampling, supports single-pass selection, and operates without an initial labeled set. Its regression framing differs from LeaDQ’s stream-query and multi-agent-RL framing. PMLR proceedings page |
| Online active learning | A broader learning setting centered on continually arriving data streams. | Selects observations from the stream for labeling, with reducing labeled-data collection costs as a key motivation. Springer Nature survey |
When LeaDQ is relevant
LeaDQ is most relevant to a system where data are distributed across clients, arrive as an unlabeled stream, and must be selectively annotated to support a common model. When evaluating whether its approach fits a particular project, check these dimensions:
- Data availability: Are examples arriving over time, or is the system selecting from a fixed unlabeled pool?
- Task: Does the application resemble the image and text tasks in LeaDQ’s reported simulations, or does it require a different problem framing such as regression?
- Selection coordination: Should clients select locally, rely on global or local selectors, or learn coordinated policies?
- Evidence needed: Is simulation evidence sufficient for the decision, or does the deployment require evaluation under its own client distributions, annotation process, and operational constraints?
These distinctions help separate LeaDQ’s proposal from other active-learning methods and clarify what would need to be tested in a specific federated system.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
