FedX: Federated Algorithms & Query Processing

Updated 3 July 2026

FedX is a set of versatile frameworks that enable efficient federated query processing and learning, addressing privacy, resource, and communication challenges.
It leverages techniques like exclusive grouping, fragment containment-aware routing, mobile transporter-based updates, and explanation-guided pruning to improve performance.
Empirical results across Linked Data queries, IoT federated learning, and unsupervised setups demonstrate significant reductions in latency, computation, and communication overhead.

FedX refers to a range of algorithmic frameworks and systems spanning federated learning, federated hyperparameter optimization, federated SPARQL query processing, communication-efficient FL, unsupervised FL, model decomposition and quantization for IoT devices, and explanation-guided network pruning. The term is used for both query federation mediators in Linked Data (FedX, originating c. 2012), and for a variety of federated learning (FL) algorithms (FedX, FedEx, etc.) developed over the past decade, targeting diverse resource, privacy, and communication bottlenecks.

Below, the primary instantiations and their research contexts are detailed.

1. Federated Query Processing: FedX as SPARQL Federation Mediator

FedX, originally introduced by Schwarte et al., is a high-performance federation mediator for Linked Data query processing. It enables transparent SPARQL querying over multiple autonomous endpoints, treating data as if it were a single integrated RDF graph. The classic FedX pipeline comprises query parsing, source selection via endpoint ASK probes, exclusive group formation (grouping patterns assignable to a singleton endpoint), join reordering using source cardinality heuristics, subquery dispatch, and result assembly (Schwarte et al., 2012).

FedX’s source selection algorithm is based on parallel ASK queries to determine which triple patterns can be answered by which endpoints. Optimization techniques include exclusive grouping (bundling patterns mapped to single endpoints), rule-based join ordering (minimize size of intermediates), and persistent caching of source selection results. Experimental results on large-scale Bio2RDF federations (29 datasets, ~4.1B triples) demonstrate sub-second evaluation for most queries in controlled (local) settings, with network round-trip times and subquery counts being the main bottlenecks in hybrid (WAN) deployments. Caching of ASK results and strategic endpoint partitioning ameliorate network-induced latencies.

2. Source-Selection Optimization: Integration of FEDRA into FedX

FedX was extended with FEDRA, a source selection framework exploiting replicated data fragments to optimize federated query routing (Montoya et al., 2015). The FEDRA extension replaces vanilla FedX’s default probe-per-pattern logic with a fragment containment-aware approach. Endpoint fragment metadata is leveraged to infer containment and coverage, grouping endpoints into equivalence classes and ensuring that for each triple pattern the minimal necessary subset of endpoints is contacted while favoring consumer-side caches over public endpoints.

FEDRA’s minimization-based set covering within Basic Graph Patterns (BGPs) enables co-location of joins on endpoints, drastically reducing numbers of public endpoint invocations (NSPS) and the overall number of sources (NSS) per query. Empirical benchmarks show that slotting FEDRA into FedX yields order-of-magnitude reductions in intermediate result volume and up to 20× execution time speedups, while answer recall remains near-perfect.

3. Federated Learning via Indirect Communication: “FedEx” with Mobile Transporters

The “FedEx” framework addresses FL in environments where direct communication between server and clients is infeasible, deploying mobile transporters (e.g., UAVs) as physical couriers of model updates (Bian et al., 2023, Bian et al., 2023). Two modes are studied:

FedEx-Sync: Synchronizes K transporters on rounds of length $\Delta = \max_k \Delta_k$ (the slowest round-trip among all transporters/routes). Each client downloads the latest global model, performs $\Delta$ local steps, and uploads cumulative updates when collected by the transporter.
FedEx-Async: Each transporter operates independently on its own round-trip schedule, with updates asynchronously aggregated by the server upon arrival.

The assignment of clients to transporters and route planning are formulated as a bi-level optimization, minimizing either the maximal round trip (Sync) or the sum-of-weighted-squares delays (Async), subject to per-transporter energy constraints. Theoretical analysis establishes that (under standard smoothness and bounded-variance assumptions) both algorithms achieve the classic $O(1/\sqrt{NT})$ convergence rate of distributed SGD, with the delay terms entering only through the assignment objective. Empirical simulation on FMNIST and SVHN shows that FedEx-Async performs best under IID, while FedEx-Sync can be more robust to data partition skew.

4. Communication-Efficient FL: Explanation-Guided Pruning (FedX)

FedX for communication-efficient FL in remote sensing employs explanation-guided pruning at the server to reduce model parameter transmission without significant loss of accuracy (Büyüktaş et al., 8 Aug 2025). After initial rounds, the server computes backpropagation-based relevance scores over a reference dataset, applies layer-wise quantile thresholding to prune the least informative filters/neurons, and transmits the resulting masked model. Clients retain their standard FL pipeline, applying the mask to maintain structural consistency.

On BigEarthNet-S2 (multi-label) and EuroSAT (single-label), FedX trims communication by up to 45%, with performance preserved within 1–4% of the unpruned baseline for moderate pruning. The technique generalizes across CNN and ViT architectures and, in some cases, regularizes the model leading to slight accuracy gains at mild sparsity. Compared to random and prior pruning baselines, explanation-guided masks retain markedly higher performance at equivalent sparsity levels.

5. Adaptive Model Decomposition and Quantization: FedX for IoT FL

In resource-constrained IoT federations, FedX implements adaptive model decomposition and server-side quantization (Lai et al., 17 Apr 2025). The global “super-net” is partitioned into nested sub-networks (e.g., via CompOFA), with each device assigned a sub-network and bit-width ( $q_u$ ) maximizing utility under its on-device time budget ( $\mu_u$ ). All quantization is performed server-side using QSGD; devices receive quantized weights but train in full precision, imposing zero quantization overhead on clients.

The bi-level optimization minimizes the sum of quantized local losses and a regularization term enforcing server-public data alignment and model consistency. Empirical results demonstrate up to 8.43× quantization time speedup, 1.5× on-device computation reductions, and 1.36× end-to-end training acceleration over FedAvg and other adaptive quantized FL baselines, with superior test accuracy even on non-IID data.

6. Expediting FL on Heterogeneous Mobile Devices: FedEx with Overlapping and Participant Selection

FedEx, targeting FL latency minimization under mobile device heterogeneity, introduces an “overlapping” protocol wherein local gradient uploads are hidden behind continued device-local computation, subject to a staleness ceiling (Geng et al., 2024). This capping ensures that each client accumulates at most $K$ extra local steps, bounding memory and model drift. Participant selection (PS) is redefined to integrate overlapping benefits, favoring clients likely to contribute low-latency updates.

The protocol incorporates a staleness-trigger metric (based on centered-kernel alignment between local and global models) to avoid overlapping until client models become sufficiently similar, and holistic PS mechanisms to avoid straggler domination. Extensive testbed results indicate speedups of 1.8–3.0× over conventional FL and overlap-based baselines, with controlled memory usage and robust performance across device and data heterogeneity.

7. Unsupervised Federated Learning: FedX with Cross Knowledge Distillation

FedX in the context of unsupervised FL implements a two-sided knowledge distillation mechanism combining self-distillation (within-client, contrastive and relational losses) and cross-client distillation from the server-side global encoder (Han et al., 2022). Each client alternates between maximizing mutual information between augmentations (local contrastive) and aligning with global prototypes (via a prediction head, global contrastive and relational). The combined loss is the unweighted sum of local/global contrastive and relational components.

Applied as a drop-in local objective for existing unsupervised FL methods (e.g., FedSimCLR, FedMoCo, FedBYOL), FedX yields 1.58–5.52pp linear-probe top-1 accuracy improvements on standard image benchmarks. Loss ablations confirm the necessity of all four components for maximum gain, and embedding analyses reveal increased angular separation of class prototypes, mitigating the effects of client drift and representation collapse.

Table: FedX—Representative Algorithms and Systems

Domain	Core Mechanism	Key Reference
SPARQL Federation	Source selection, parallel ASK, exclusive grouping	(Schwarte et al., 2012, Montoya et al., 2015)
Indirect FL Comm.	Model transport via UAVs, route optimization	(Bian et al., 2023, Bian et al., 2023)
Comm.-Eff. FL RS	Explanation-guided pruning	(Büyüktaş et al., 8 Aug 2025)
IoT FL	Model decomposition, server-side quantization	(Lai et al., 17 Apr 2025)
FL Latency	Overlapping protocol with PS utility	(Geng et al., 2024)
Unsupervised FL	Cross knowledge distillation	(Han et al., 2022)
HPO in FL	Single-shot weight-sharing optimization	(Khodak et al., 2021)

Each instantiation addresses a specific challenge in federated computing—be it communication minimization, resource adaptation, unsupervised learning, or query answer optimization—and advances the state of the art in its domain.

FedX, in its diverse technical forms, represents a set of principled systems for scalable, efficient, and adaptive federated computation and query processing, continuously evolving to address emerging constraints in privacy, heterogeneity, and resource limitation.