FedDeCAB: Semi-Decentralized Time-Series FL

• FedDeCAB is a semi-decentralized federated learning method designed for time-series prediction in IoT deployments, integrating client availability budgets.
• It combines availability-aware client selection with neighbor-assisted parameter sharing to mitigate challenges from intermittent connectivity and resource constraints.
• Empirical results indicate that FedDeCAB accelerates convergence and reduces RMSE compared to methods like FedAvg under varied availability scenarios.

FedDeCAB is a federated learning method for time-series prediction under client availability budgets, introduced as a semi-decentralized scheme that combines availability-aware client selection at the server with neighbor-assisted parameter sharing for clients that are temporarily disconnected from the server. In the supplied literature, the same string is also mapped to several conceptually related but distinct constructs, including FedDDL for deconfounding and debiasing in out-of-distribution federated vision, Decaf for data-distribution decomposition attacks, FCaC for capability-based boundary admission, FedDA for decoupled adaptive optimization, and DecantFed for semi-synchronous client clustering and bandwidth allocation. Only "Semi-decentralized Federated Time Series Prediction with Client Availability Budgets" explicitly introduces a method named FedDeCAB (Bao et al., 3 Sep 2025, Qi et al., 8 May 2025, Dai et al., 2024, Fenoglio et al., 18 Mar 2026, Jin et al., 2022, Yu et al., 2024).

1. Nomenclature and scope

A recurring source of confusion is that “FedDeCAB” is not used uniformly across the cited papers. In the time-series setting, it is the exact method name. In the other cases, the supplied clarifications explicitly map the string to methods or architectures whose published names differ.

Usage in source material	Published name	Relation to “FedDeCAB”
(Bao et al., 3 Sep 2025)	FedDeCAB	Explicit method name
(Qi et al., 8 May 2025)	FedDDL	Conceptually aligned mapping
(Dai et al., 2024)	Decaf	Attack-level mapping
(Fenoglio et al., 18 Mar 2026)	FCaC	Capability-based admission mapping
(Jin et al., 2022)	FedDA	Decoupled adaptive optimization mapping
(Yu et al., 2024)	DecantFed	Dynamic clustering/bandwidth/workload mapping

In its explicit usage, FedDeCAB addresses a systems regime in which federated clients are constrained not only by data heterogeneity but also by limited energy and communication budgets, intermittent connectivity, and dynamic online/offline behavior. The method targets time-series prediction in IoT deployments and is formulated around a central server, a client set $C=\{1,\dots,N\}$, a sequence of global rounds $t=1,\dots,T$, local model parameters $w_i^t \in \mathbb{R}^d$, and a global model $w^t \in \mathbb{R}^d$ (Bao et al., 3 Sep 2025).

This suggests that the most precise interpretation of the term is context-dependent rather than canonical. In current usage, FedDeCAB is best treated as a named semi-decentralized time-series FL method, while other appearances in the supplied material are mappings or analogical extensions rather than original nomenclature.

2. System model and optimization objective

FedDeCAB assumes local time-series datasets $D_i$ containing trajectory slices, such as AIS vessel tracks or GPS taxi traces, and uses recurrent sequence models trained locally. The experimental configuration employs an LSTM with 128 units and 6 time-steps (Bao et al., 3 Sep 2025).

The local forecasting loss is written as

$L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$

with mean squared error as the supervised loss and RMSE used for evaluation:

$$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$

The global objective is summarized as

$$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$

where $f_i$ is the local prediction loss on client $i$, $t=1,\dots,T$0 encodes availability over rounds, and $t=1,\dots,T$1 is a neighbor-dependent term used when collaboration occurs during disconnection (Bao et al., 3 Sep 2025).

For offline-client updates, the collaborative parameter is chosen as

$t=1,\dots,T$2

and the local objective becomes

$t=1,\dots,T$3

where $t=1,\dots,T$4, and

$t=1,\dots,T$5

is applied over normalized FC-layer parameter distributions of $t=1,\dots,T$6 and $t=1,\dots,T$7. When the server is reachable, clients revert to standard local ERM without this neighbor regularizer (Bao et al., 3 Sep 2025).

Communication is budgeted. Each client $t=1,\dots,T$8 has an availability budget $t=1,\dots,T$9, a dynamic online probability $w_i^t \in \mathbb{R}^d$0, and a per-round communication cost $w_i^t \in \mathbb{R}^d$1, under the constraint

$w_i^t \in \mathbb{R}^d$2

where $w_i^t \in \mathbb{R}^d$3 denotes participation at round $w_i^t \in \mathbb{R}^d$4. FedDeCAB also tracks activeness,

$w_i^t \in \mathbb{R}^d$5

where $w_i^t \in \mathbb{R}^d$6 counts how many times client $w_i^t \in \mathbb{R}^d$7 has uploaded to the server by round $w_i^t \in \mathbb{R}^d$8 (Bao et al., 3 Sep 2025).

The architecture therefore couples statistical learning with explicit resource accounting. A plausible implication is that availability is treated as a first-class optimization variable rather than a nuisance condition.

3. Availability-aware client selection

The client-selection mechanism combines heterogeneity utility, participation fairness, and compensation for stragglers and late joiners. For each available client $w_i^t \in \mathbb{R}^d$9 at round $w^t \in \mathbb{R}^d$0, the server computes $w^t \in \mathbb{R}^d$1, the divergence between the broadcast global model and the client’s locally updated model, together with the activeness score $w^t \in \mathbb{R}^d$2. Two descending priority lists are then formed: one by KL divergence, giving position $w^t \in \mathbb{R}^d$3, and one by activeness, giving position $w^t \in \mathbb{R}^d$4 (Bao et al., 3 Sep 2025).

Early in training, FedDeCAB prioritizes high divergence through a decaying quadratic weight curve. For $w^t \in \mathbb{R}^d$5,

$w^t \in \mathbb{R}^d$6

with

$w^t \in \mathbb{R}^d$7

where $w^t \in \mathbb{R}^d$8 is the number of available clients. When $w^t \in \mathbb{R}^d$9, the method switches to the linear form

$D_i$0

This transition is described as moving from divergence-driven prioritization toward a regime that accelerates convergence later in training (Bao et al., 3 Sep 2025).

The ranking weight is

$D_i$1

where $D_i$2 is a per-client boost that decays by $D_i$3 after each upload. If

$D_i$4

then the weight is further boosted by $D_i$5, i.e., $D_i$6 (Bao et al., 3 Sep 2025).

Selection is performed by ranking available clients according to $D_i$7 and taking up to $D_i$8. The implementation in Algorithm 1 uses deterministic top-$D_i$9 selection based on sorted weights, although the paper also gives a probabilistic interpretation through normalized weights,

$L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$0

or a softmax variant (Bao et al., 3 Sep 2025).

This selection rule differs from uniform random sampling by explicitly favoring underrepresented or under-participating clients. The design intent is to reduce model bias early, then improve convergence efficiency as training progresses.

4. Semi-decentralized offline collaboration

FedDeCAB’s distinctive systems contribution is the treatment of temporary server disconnection. When a client is offline from the server but can still reach nearby peers, it may obtain partial model parameters from its nearest neighbors for joint optimization (Bao et al., 3 Sep 2025).

The peer topology is modeled as an undirected graph $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$1, with neighbor set

$L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$2

In offline rounds, which occur at a lower frequency than centralized rounds, client $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$3 requests FC-layer parameters from neighbors $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$4, forming candidate models $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$5. It evaluates the empirical loss of each candidate head using its own local data and selects

$L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$6

The subsequent local update adds the KL term $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$7, which pulls the offline model toward the best-performing neighbor head in parameter space (Bao et al., 3 Sep 2025).

Only the fully connected layers are shared in this peer-to-peer phase. In the LSTM configuration reported in the paper, the FC layers account for about $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$8 of the total model parameters, specifically $L_i(w)=\mathbb{E}_{(x,y)\sim D_i}[\ell(f_w(x),y)],$9 of $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$0. The ratio is written as

$$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$1

corresponding to an approximately $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$2 reduction in payload relative to full-model sharing (Bao et al., 3 Sep 2025).

The server-side aggregation rule remains standard FedAvg:

$$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$3

where $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$4 is the selected client set and $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$5 is the local sample count used by client $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$6 (Bao et al., 3 Sep 2025).

This yields a hybrid design: centralized aggregation when connectivity exists, and lightweight semi-decentralized cooperation when it does not. The paper notes that one could generalize the offline mechanism to weighted neighbor interpolation, but the implemented method uses argmin selection over empirical loss rather than convex interpolation (Bao et al., 3 Sep 2025).

5. Availability scenarios, training protocol, and empirical behavior

FedDeCAB is evaluated on two real-world trajectory domains: vessel trajectory data from China sea AIS and the T-Drive taxi trajectory sample. Partitions are constructed both by equal allocation and by object identity, such as one vessel per client or four vehicles per client, to induce non-i.i.d. skew (Bao et al., 3 Sep 2025).

Three availability scenarios are modeled. In random trajectory availability, per-point availability is drawn independently from a Dirichlet distribution. In region-based availability, points in weak-signal regions use $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$7 and others use $$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$8. In data-size-based availability, larger-data clients are assigned higher availability than smaller-data clients by a size threshold. Client online/offline behavior follows

$$RMSE(y,\hat{y})=\sqrt{\frac{1}{n}\sum_{k=1}^n (y_k-\hat{y}_k)^2}.$$9

These dynamics affect both centralized client selection and offline neighbor sharing (Bao et al., 3 Sep 2025).

The reported training setup uses an LSTM with 128 units, input sequence length $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$0, batch size $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$1, learning rate $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$2, local epochs $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$3, sampling ratio $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$4, equal client budgets of $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$5 total server uploads, offline rounds at $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$6 of central frequency, and FC-only neighbor sharing with $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$7 parameters (Bao et al., 3 Sep 2025).

The baselines are FedAvg, FedProx, MOON, and FedCAB, together with FedProx+ and MOON+, which add FedCAB-style ranking while preserving the original local objectives (Bao et al., 3 Sep 2025).

The results are summarized qualitatively rather than through tabulated scalar RMSE values in the text. Across random, regional, and data-size-based availability scenarios for both datasets, FedAvg, FedProx, and MOON with uniform random sampling exhibit high volatility and inferior RMSE under constrained availability and intermittent connectivity. FedCAB improves selection under availability budgets but cannot utilize offline clients’ knowledge, and its convergence is slower. FedDeCAB, which combines ranking-based selection with semi-decentralized neighbor collaboration, consistently achieves faster convergence and lower RMSE under high offline rates and limited budgets. The paper also reports that FedProx+ and MOON+ outperform their original variants, highlighting the value of availability-aware selection, and emphasizes the approximately $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$8 reduction in peer-to-peer communication payload achieved by FC-only neighbor sharing (Bao et al., 3 Sep 2025).

The empirical narrative places FedDeCAB in a design space where the principal bottleneck is not only statistical heterogeneity but also intermittent participation. Its performance claims are therefore tied as much to scheduling and communication design as to the local learning objective.

6. Limitations, assumptions, and related interpretations

The paper does not provide formal convergence theorems for FedDeCAB. The discussion instead gives an intuitive rationale: early-stage quadratic weighting favors high-KL clients to reduce bias from underrepresented distributions, late-stage linear weighting accelerates convergence, and offline neighbor rounds add a proximal-like KL regularizer that stabilizes disconnected clients (Bao et al., 3 Sep 2025). This suggests a hybrid between selection-aware FL and lightweight graph-assisted personalization, but the text explicitly identifies rigorous convergence analysis under dynamic availability as future work.

Several limitations are stated directly. If neighbors have poor or biased heads, offline clients may be pulled toward suboptimal directions, although the empirical-loss-based choice of $$f(w):=\frac{1}{N}\sum_{i=1}^N [f_i(w,D_i,\xi_i,\tau_i)+g_i(w,v_i^*)],$$9 is intended to mitigate this. Transmitting FC-layer parameters may leak information, so additional privacy mechanisms such as secure aggregation or differential privacy on transmitted heads are possible, albeit with utility or communication trade-offs. Deterministic top-$f_i$0 ranking may reduce randomness and exploration relative to probabilistic sampling. Robustness may also degrade under extreme offline rates or sparse neighbor graphs, and future designs may need adaptive neighborhood expansion or relaying (Bao et al., 3 Sep 2025).

The broader interpretive landscape is unusually crowded. In the supplied clarifications, “FedDeCAB” is also used as a label for at least five other ideas: causal attribute debiasing through counterfactual backgrounds and causal prototypes in FedDDL (Qi et al., 8 May 2025); full class-distribution reconstruction in the passive privacy attack Decaf (Dai et al., 2024); decentralized capability-based admission in FCaC through the KYO $f_i$1 ECT $f_i$2 PoP trust chain (Fenoglio et al., 18 Mar 2026); decoupled adaptive optimization in FedDA, where global momentum is exactly decomposed across clients (Jin et al., 2022); and semi-synchronous client clustering, bandwidth allocation, and workload optimization in DecantFed (Yu et al., 2024).

Accordingly, the main misconception to avoid is treating FedDeCAB as a stable umbrella term with a single accepted expansion across federated learning. The source material supports a narrower conclusion: FedDeCAB is explicitly a semi-decentralized availability-aware time-series FL method in (Bao et al., 3 Sep 2025), while its appearance elsewhere is interpretive, analogical, or terminological mapping rather than original paper nomenclature.