FedEPD: Context-Dependent Federated Learning

• FedEPD is a context-dependent term that encompasses distinct models for energy demand prediction, graph learning via dual decoupling, and fall-detection in wearable healthcare.
• Its methodologies include variational inference with expectation propagation, synchronous gradient aggregation, and energy-guided topological purification.
• Clarifying the terminological scope of FedEPD is crucial to avoid confusion with similar acronyms like FedEP and unrelated usage in other domains.

FedEPD is not a single stable designation in the recent arXiv literature. Depending on context, it may refer to a shorthand for federated energy demand prediction in electric-vehicle charging networks, a distinct framework for federated long-tailed graph learning, or an alternative label for a federated fall-detection pipeline. In an influential probabilistic federated learning paper, however, the acronym does not appear at all: the method is “FedEP,” not “FedEPD” (Guo et al., 2023). The term therefore requires disambiguation before any technical discussion.

1. Terminological scope

Across the cited literature, “FedEPD” has multiple, non-equivalent uses, and one acronym collision is entirely unrelated to federated learning.

Usage	Paper	Defining characterization
Not defined; correct term is FedEP	(Guo et al., 2023)	“FedEPD” does not occur; the method is Federated Expectation Propagation
Conceptual label for FEDL	(Saputra et al., 2019)	Federated energy demand learning for EV charging stations
Formal method name FedEPD	(Guo et al., 23 Jun 2026)	Energy-guided dual decoupling for federated long-tailed graph learning
Alternative label for EPFL	(Rao et al., 23 Oct 2025)	Ensembled penalized federated learning for falling people detection
Unrelated acronym EPD	(Shao et al., 2020)	STAR Event Plane Detector in relativistic heavy-ion experiments

The most important clarification is terminological. In “Federated Learning as Variational Inference: A Scalable Expectation Propagation Approach,” the paper explicitly states that the acronym “FedEPD” does not occur and is not defined; the actual algorithm is “FedEP,” with a damped update parameter $\delta$, while $D$ denotes either the dataset partitioning or a divergence such as KL (Guo et al., 2023). By contrast, “Towards Federated Long-Tailed Graph Learning: An Energy-Guided Dual Decoupling Approach” introduces a method explicitly named FedEPD (Guo et al., 23 Jun 2026).

A further source of confusion is acronym overlap outside federated learning. In the STAR instrumentation literature, EPD denotes the Event Plane Detector, a plastic scintillator–based detector installed in the forward rapidity region of STAR; this usage is unrelated to any federated algorithmic framework (Shao et al., 2020).

2. FedEP in variational federated learning

The probabilistic method most likely to be mistaken for “FedEPD” is FedEP, which formulates federated learning as distributed posterior inference rather than distributed loss minimization. The Bayesian setup assumes client-partitioned data $D=\bigcup_i D_i$, a prior $p(w)$, and conditionally independent client likelihoods $p(D_i\mid w)$, so that the exact posterior factorizes as

$p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$

Because exact inference is intractable, the paper adopts a variational inference perspective and seeks a global variational posterior $q(w)\in Q$ maximizing the ELBO,

$$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$

Its concrete variational family is a mean-field Gaussian with diagonal covariance, parameterized in both moment and natural forms, with Gaussian product and quotient identities enabling closed-form manipulation of sites, cavities, and global approximations (Guo et al., 2023).

Expectation propagation introduces per-client site terms $t_i(w)$ so that

$q(w)\propto p(w)\prod_{i=1}^N t_i(w),$

with cavity distribution $D$0 and tilted distribution $D$1. In natural parameters, the EP updates are additive over the prior and site factors, and the paper uses damping $D$2 to stabilize parallel updates. The implementation reinterprets the natural-parameter increments $D$3 and $D$4 as gradients, which allows adaptive optimizers such as Adam, Adagrad, and SGD with momentum to be applied on both client and server (Guo et al., 2023).

Several scalable tilted-distribution approximations are developed. SG-MCMC with moment matching adds a cavity regularizer to the local negative log-likelihood and keeps client compute essentially the same order as FedAvg. A scaled identity covariance approximation sets $D$5, halving communication because no covariance vector is sent. Two curvature-aware alternatives are also provided: a Laplace approximation with diagonal Fisher and NGVI with diagonal Fisher. To eliminate per-client state, the paper further introduces a stateless stochastic EP variant, FedSEP, in which the global approximation is $D$6 times $D$7 copies of a single shared factor, making it more memory-efficient under extreme partial participation (Guo et al., 2023).

The federated message-passing protocol is server–client EP. The server broadcasts the current global Gaussian $D$8 to a sampled subset of clients $D$9; each client forms its cavity, approximates the tilted distribution by one of the scalable methods, and sends site-update messages $D=\bigcup_i D_i$0 to the server; the server then updates the global natural parameters with damping and an adaptive optimizer. With diagonal Gaussians, a client message comprises two $D=\bigcup_i D_i$1-dimensional vectors, while scaled identity covariance requires only the mean update and therefore halves the payload (Guo et al., 2023).

Empirically, FedEP and FedSEP outperform strong baselines on CIFAR-100, StackOverflow tag prediction, and EMNIST-62. On CIFAR-100 with a ResNet-18 of 11.2M parameters, FedAvg reaches $D=\bigcup_i D_i$2 test accuracy by round 1500, whereas FedEP(I, scaled identity) reaches $D=\bigcup_i D_i$3 and FedEP(M, SG-MCMC+moments) reaches $D=\bigcup_i D_i$4; the rounds to $D=\bigcup_i D_i$5 accuracy drop from $D=\bigcup_i D_i$6 for FedAvg to $D=\bigcup_i D_i$7 for FedEP(I) and $D=\bigcup_i D_i$8 for FedEP(M). On StackOverflow, stateless SEP improves micro-F1 from $D=\bigcup_i D_i$9 for FedAvg to $p(w)$0 for FedSEP(V). On CIFAR-100 calibration, FedAvg has ECE $p(w)$1, while FedEP(V, NGVI) reaches ECE $p(w)$2 for the point estimate and $p(w)$3 for the marginalized posterior (Guo et al., 2023).

3. FedEPD as federated energy demand prediction

In electric-vehicle charging networks, “FedEPD” is used as a conceptual label for the paper’s Federated Energy Demand Learning framework, FEDL. The task is next-period energy demand prediction at charging stations, using transaction logs whose features include CS ID, day-of-week, and hour-of-day, with consumed energy as the learning target. The supervised objective is mean squared error,

$p(w)$4

and evaluation is by RMSE on test data (Saputra et al., 2019).

The system consists of a Charging Station Provider and multiple charging stations. In centralized EDL, all stations send raw transaction logs to the provider, which trains a deep neural network centrally. In FEDL, each charging station trains locally and sends gradients or model updates instead of raw data. The aggregation rule in the paper is synchronous gradient averaging,

$p(w)$5

after which the provider updates the global model with Adam and broadcasts it back to all stations. The paper explicitly uses synchronous aggregation so that the provider waits for all local gradients each round, ensuring zero gradient staleness (Saputra et al., 2019).

The predictor is a DNN with two hidden layers of 64 neurons each, tanh activations, and dropout fraction $p(w)$6 after the last hidden layer. Adam is used with initial step size $p(w)$7. To address bias from heterogeneous station behavior, the paper also introduces clustering-based EDL using constrained K-means on charging-station locations, with min/max cluster-size constraints and per-cluster training thereafter (Saputra et al., 2019).

The Dundee City, UK dataset contains 65,601 transactions across 58 charging stations. At the 80% training ratio, the reported RMSE values are 7.18 for KNR, 6.57 for MLPR, 6.55 for SGDR, 6.47 for DT, 6.46 for SVR, 6.35 for RF, 5.86 for centralized EDL, 5.81 for FEDL, 5.77 for EDL + Clustering, and 5.76 for FEDL + Clustering. The paper reports up to $p(w)$8 lower RMSE for FEDL + Clustering versus baseline methods at the 80% training ratio, and communication-overhead reduction by up to $p(w)$9 relative to centralized methods because raw datasets are not uploaded (Saputra et al., 2019).

In this literature stream, then, FedEPD is not the authors’ formal algorithmic name but a domain-specific shorthand for federated energy demand prediction operationalized by FEDL. The core contribution is not probabilistic message passing or graph purification, but synchronous federated regression over charging-station logs, optionally refined by location-based clustering (Saputra et al., 2019).

4. FedEPD in federated long-tailed graph learning

A formally named FedEPD appears in federated graph learning as an energy-guided dual decoupling framework for long-tailed, non-IID, and heterophilic settings. The global graph $p(D_i\mid w)$0 is partitioned across $p(D_i\mid w)$1 clients into local subgraphs $p(D_i\mid w)$2, with long-tailed global class counts $p(D_i\mid w)$3 and a large imbalance ratio $p(D_i\mid w)$4. The paper’s motivation is twofold: majority classes dominate the global model, and minority nodes are structurally submerged in heterophilic, head-dominated neighborhoods, so standard GNN message passing propagates dominant-class features into minority representations (Guo et al., 23 Jun 2026).

FedEPD separates “where noise lives” from “what we must calibrate” through a dual decoupling paradigm. The first branch is topological purification by distribution-aware Dirichlet energy pruning. For a node signal $p(D_i\mid w)$5, the Dirichlet energy is

$p(D_i\mid w)$6

FedEPD avoids eigendecomposition and uses an edge-level discrepancy proxy based on degree-normalized feature discrepancy plus semantic inconsistency $p(D_i\mid w)$7. After standardization and empirical-CDF mapping, the fused score is

$p(D_i\mid w)$8

with

$p(D_i\mid w)$9

and edges above the client-specific quantile threshold are pruned:

$p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$0

This confines message passing to structurally consistent neighborhoods and is designed to protect minority nodes from dominant-class leakage (Guo et al., 23 Jun 2026).

The second branch is semantic recalibration by server-assisted global prototypes. For each class, the client computes Personalized PageRank centrality with restarts from the class’s labeled training nodes,

$p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$1

using $p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$2 and $p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$3 steps, then selects $p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$4 elite nodes with $p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$5. Their encoded features over the purified topology are summed into local prototype components $p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$6, and the server aggregates

$p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$7

These prototypes are reintroduced by spatial low-pass prototype injection. With $p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$8,

$p(w\mid D)\propto p(w)\prod_{i=1}^N p(D_i\mid w).$9

and for a labeled node $q(w)\in Q$0,

$q(w)\in Q$1

Only the low-frequency component receives the prototype, while high-frequency node-specific detail is preserved (Guo et al., 23 Jun 2026).

Optimization alternates between two stages. Stage 1 trains encoder and classifier jointly on the purified graph using

$q(w)\in Q$2

Stage 2 freezes the encoder, injects prototypes, and applies topology-aware logit adjustment. The homophily gate for class $q(w)\in Q$3 is formed from node-level homophily estimates

$q(w)\in Q$4

leading to

$q(w)\in Q$5

and the class margin

$q(w)\in Q$6

Adjusted logits $q(w)\in Q$7 are optimized with a calibration loss $q(w)\in Q$8, while the server aggregates model parameters by FedAvg and prototypes by mean-of-sums (Guo et al., 23 Jun 2026).

Experiments use a 2-layer GCN with hidden size 64, 200 rounds, 3 local epochs per round, and $q(w)\in Q$9 clients on CoraFull, ogbn-arxiv, Amazon-Electronics, Amazon-Clothing, Roman-Empire, and Email. FedEPD achieves state-of-the-art performance across six datasets, with absolute improvements of up to $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$0 in Accuracy and $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$1 in Macro-F1. On Amazon-Electronics, it attains Macro-F1 $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$2 and balanced accuracy $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$3. Ablations show that removing topological purification sharply degrades Macro-F1, for example on Roman-Empire from $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$4 to $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$5 and on Amazon-Electronics from $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$6 to $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$7, while removing global consensus or decoupled recalibration also causes consistent deterioration (Guo et al., 23 Jun 2026).

5. FedEPD as EPFL in falling people detection

In wearable-sensor healthcare, the name “FedEPD” is used as an alternative label for the paper’s EPFL framework, “Ensembled Penalized Federated Learning.” The task is falling people detection using multivariate time series from sensors placed on the ankle, chest, and belt, aligned into a 9-dimensional sequence and segmented with sliding windows of size 20 and stride 1. The authors state that EPFL can be equivalently referred to as FedEPD in this context (Rao et al., 23 Oct 2025).

The local model is a two-layer LSTM with hidden size 128, followed by batch normalization, ReLU, a fully connected layer, and a sigmoid output. Local training minimizes a FedProx-style penalized objective

$$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$8

where $$\mathrm{ELBO}(q)=\mathbb{E}_q[\log p(w)] + \sum_{i=1}^N \mathbb{E}_q[\log p(D_i\mid w)] - \mathbb{E}_q[\log q(w)].$$9 is binary cross-entropy, $t_i(w)$0, and $t_i(w)$1. Adam is used with learning rate $t_i(w)$2, batch size 32, and $t_i(w)$3 local epochs per round (Rao et al., 23 Oct 2025).

Server aggregation uses Specialized Weighted Aggregation. First, FedNova-style normalization rescales local updates:

$t_i(w)$4

Second, a trimmed mean removes $t_i(w)$5 extreme values at both ends per parameter dimension, with $t_i(w)$6. Third, exponential moving average fusion updates the global model:

$t_i(w)$7

with $t_i(w)$8. For deployment, each client ensembles its personal model and the global model:

$t_i(w)$9

and triggers an alert if the ensemble probability exceeds $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$0 (Rao et al., 23 Oct 2025).

The framework also includes homomorphic encryption in transit through TenSEAL and CKKS. Training and inference are still executed in plaintext locally and on the server because ciphertext computation is too costly in the current implementation. Continual learning is incorporated through human-in-the-loop feedback: a confirmed or refuted alert $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$1 is appended to the local dataset and reused in later federated rounds (Rao et al., 23 Oct 2025).

On the LDPA dataset, EPFL+SWA achieves Accuracy $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$2, Precision $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$3, Recall $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$4, and F1 $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$5. The centralized LSTM baseline reaches Accuracy $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$6, Precision $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$7, Recall $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$8, and F1 $q(w)\propto p(w)\prod_{i=1}^N t_i(w),$9. FedAvg shows Recall $D$00 but Precision $D$01 and F1 $D$02, indicating severe false positives, while HBOS under the merged-sensor setting reaches Recall $D$03 but Precision $D$04 and F1 $D$05. The per-client recall for EPFL+SWA is 0.8444 for A, 1.0000 for B, 0.8222 for C, 0.9114 for D, and 0.8478 for E (Rao et al., 23 Oct 2025).

6. Related names, distinctions, and recurring misconceptions

The literature supports three direct corrections to common misunderstandings. First, “FedEPD” should not be used as a synonym for the expectation-propagation method in federated variational inference; that paper defines and studies “FedEP” and “FedSEP,” and explicitly states that “FedEPD” does not occur and is not defined (Guo et al., 2023). Second, in application-specific work the acronym may be a descriptive shorthand rather than the authors’ formal method name: the EV paper names its method FEDL, and the fall-detection paper names its method EPFL, even though both summaries map them to “FedEPD” in context (Saputra et al., 2019, Rao et al., 23 Oct 2025). Third, a standalone algorithm actually named FedEPD appears in federated long-tailed graph learning, where the acronym denotes an energy-guided dual decoupling approach rather than expectation propagation, energy demand learning, or fall detection (Guo et al., 23 Jun 2026).

A further distinction is required between FedEPD and methods named FedEP. Besides the variational-inference FedEP, there is also a decentralized federated learning algorithm called FedEP that uses Gaussian Mixture Models to fit local data distributions, shares only GMM parameters and sample counts, estimates a neighborhood or global distribution as a mixture-of-mixtures, and computes aggregation weights by normalized KL divergences between pooled and local distributions. That method is aimed at decentralized aggregation under non-IID data and is unrelated to both long-tailed graph FedEPD and the EV or fall-detection usages of the acronym (Feng et al., 2024).

Taken together, these works show that “FedEPD” is best understood as a context-dependent label rather than a single canonical method family. In one strand of work it is a misreading of FedEP; in two application papers it serves as a shorthand for domain-specific federated prediction systems; and in federated graph learning it is the formal name of a specific dual-decoupling algorithm grounded in Dirichlet energy pruning, prototype consensus, and alternating optimization (Guo et al., 2023, Saputra et al., 2019, Rao et al., 23 Oct 2025, Guo et al., 23 Jun 2026).