Active Client Selection in Federated Trajectory Prediction with Uncertainty-Awareness and Heterogeneous Complexity

Yiming Xie; Muzi Peng; Fei Miao; Ningfang Mi; Lili Su

cs.LGAug 27, 2026

Active Client Selection in Federated Trajectory Prediction with Uncertainty-Awareness and Heterogeneous Complexity

Authors: Yiming Xie, Muzi Peng, Fei Miao, Ningfang Mi, Lili Su

Organizations: Northeastern University, Boston, MA, USA · University of Connecticut, Storrs, CT, USA

Abstract

Training sequence models such as transformers is now standard for autonomous vehicle trajectory prediction, yet assembling high-quality centralized datasets remains challenging because real-world trajectories are fragmented across regions and vehicles. Federated Learning (FL) offers a natural alternative, but faces two distinctive challenges: high scene uncertainty arising from trajectory or map ambiguity, and cross-scene complexity heterogeneity caused by diverse map topology, traffic density, agent composition, and driving behaviors. We propose a family of active client selection methods that progressively incorporate awareness of scene uncertainty and complexity to prioritize informative clients. Our uncertainty-aware selectors use per-client negative log-likelihood under an uncertainty-aware global objective and estimated aleatoric uncertainty. We further develop a selector that jointly considers scene complexity and uncertainty, motivated by the intuition that knowledge from complex scenes can transfer to easier ones. Experiments on Argoverse show that federated trajectory prediction outperforms locally trained models. Uncertainty-aware selection accelerates convergence and improves minADE, minFDE, and MR. Under strong scene-complexity heterogeneity, our joint complexity- and uncertainty-aware selector achieves the best generalization and further accelerates convergence, demonstrating the benefit of prioritizing complex and informative scenes.

Figures & tables

Fig. 1 : Uncertainty and heterogeneity in FL for autonomous driving. Left: uncertainty in individual traffic scenes—the past trajectory of a vehicle may admit multiple valid futures and exhibit higher uncertainty in OOD scenes. Right: scene complexity heterogeneity–arising from different map and traffic pattern distributions (urban, suburban, highway).

Fig. 2 : Evolution of the gradient error norm $\|e_{r}\|=\|g_{r}-\nabla F(w_{r})\|$ over communication rounds.The decreasing trend indicates that the discrepancy between the selected-client update and the full gradient gradually diminishes during training.

Method	Description
Random	Clients are selected randomly at each training round.
High Loss	The top $f_{1}$ fraction (30%) of clients with the highest training cross-entropy loss and Negative Log-Likelihood (NLL) are selected.
FLTP-NLL	Clients are selected based solely on Negative Log-Likelihood (NLL), without explicitly considering rank or loss separately.
High Rank + Random	The top-ranked $\nu$ clients are selected based on complexity scores, and the remaining $(f_{1}C-\nu)$ clients are selected randomly.
FLTP-DR	Clients are selected using a configurable parameter $\nu$ balancing between high-rank (trajectory complexity) and high-loss criteria, totaling $f_{1}=30\%$ of all clients.

TABLE I : Summary of baseline methods used in evaluation.

Panel of Figure 4 — Fig. 3 : Round-wise comparison between FLTP and the local model of client 0 without FL. Fraction of clients selected for communication in each round $f_{1}$ is set to be 0.1.

Panel of Figure 5 — Fig. 4 : Round-wise comparison between FLTP and FLTP-NLL

Panel of Figure 6 — Fig. 5 : Round-wise comparison between FLTP and FLTP-AU

Model	Round	NLL( $\downarrow$ )	minADE( $\downarrow$ )	minFDE( $\downarrow$ )	MR( $\downarrow$ )
Centralized HiVT [ 6 ]	-	0.467	0.685	1.028	0.104
Client 0 w/o FL	50	0.897	0.992	1.732	0.207
FLTP	50	0.632	0.818	1.321	0.143
FLTP-NLL( $f_{2}$ =0.15)	50	0.634	0.821	1.327	0.141
FLTP-NLL( $f_{2}$ =0.30)	50	0.637	0.819	1.321	0.143
FLTP-AU( $f_{2}$ =0.15)	50	0.632	0.819	1.325	0.141

TABLE II : Performance on Argoverse Validation Set. Fraction of clients selected for communication in each round

f_{1}

is set to be 0.1.

Method	$\alpha$	minADE	minFDE	Miss Rate (MR)
Random	0.1	0.754	1.310	0.121
High Loss	0.1	0.703	1.080	0.115
FLTP-NLL	0.1	0.712	1.140	0.101
High Rank + Random	0.1	0.691	1.010	0.103
FLTP-DR	0.1	0.669	0.910	0.101
Random	0.5	0.722	1.210	0.121

TABLE III : Performance comparison across selection methods (selection ratio

f_{1}=30\%

Panel of Figure 9 — Fig. 6 : Convergence behavior under $\alpha=0.1$ .

Panel of Figure 10 — Fig. 7 : Convergence behavior under $\alpha=0.5$ .

Panel of Figure 11 — Fig. 8 : Convergence behavior under $\alpha=1$ .

Panel A: Training mixture $\mathsf{Mix}(p_{H},p_{M},p_{E})$
Panel	Setting	minADE	minFDE	MR
	$\mathsf{Mix}(0.75,0,0.25)$	0.698	1.098	0.115
	$\mathsf{Mix}(0.75,0.25,0)$	0.697	1.096	0.114
	$\mathsf{Mix}(0.50,0,0.50)$	0.721	1.110	0.118
	$\mathsf{Mix}(0.50,0.50,0)$	0.710	1.097	0.114
	$\mathsf{Mix}(0.25,0,0.75)$	0.738	1.210	0.119

TABLE IV : Consolidated results. Panel A: impact of training mixture

\mathsf{Mix}(p_{H},p_{M},p_{E})

on a fixed mixed- complexity validation set

\mathcal{D}_{\mathrm{val}}^{\mathrm{mix}}

. Panel B: effect of selection ratio

f_{1}

. Panel C: effect of balance parameter

n

in FLTP-DR.

Fig. 9 : Impact of training and validation complexity on forecasting metrics. Training with hard workloads reduces minADE/minFDE, indicating that exposure to complex scenarios yields more robust representations. Evaluating on hard validation data is most demanding (and thus most discriminative), while performance remains strong on easy / moderate sets, evidencing effective generalization.

Yiming Xie is a Ph.D. student in the Department of Electrical and Computer Engineering at Northeastern University, Boston, Massachusetts. His research focuses on federated learning optimization schemes, data loader allocation for distributed deep learning, and NVMe over Fabrics SSDs.

Table 14

Muzi Peng was a Ph.D. student at Northeastern University, majoring in Electrical and Computer Engineering. He was primarily working on autonomous vehicle systems, with a particular focus on surrounding perception. He graduated with a B.S. in 2022 from Nanjing University.

Table 15

Fei Miao is the Pratt & Whitney Associate Professor of the School of Computing and a Courtesy Faculty of the Department of Electrical and Computer Engineering at the University of Connecticut. She received her Ph.D. degree and the Best Doctoral Dissertation Award in Electrical and Systems Engineering, with a dual M.S. degree in Statistics from the University of Pennsylvania in 2016. She was a postdoctoral researcher at the GRASP Lab and the PRECISE Lab of UPenn from 2016 to 2017.

Table 16

Ningfang Mi is an Associate Professor in the Department of Electrical and Computer Engineering at Northeastern University. She received her Ph.D. in Computer Science from the College of William and Mary in 2009 under the guidance of Prof. Evgenia Smirni, her M.S. in Computer Science from the University of Texas at Dallas in 2004, and her B.S. in Computer Science from Nanjing University in 2000.

Table 17

Lili Su is an Assistant Professor in the Department of Electrical and Computer Engineering at Northeastern University. She received her M.Sc. (2014) and Ph.D. (2017) in Electrical and Computer Engineering from the University of Illinois at Urbana–Champaign (UIUC). Prior to joining Northeastern, she was a postdoctoral researcher at the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT.

Table 18

Explore similar work

Choose Wisely and Privately: Proactive Client Selection for Fair and Efficient Federated Learning
May 20, 2026Adda Akram Bendoukha, Heber Hwang Arcolezi, Nesrine Kaaniche +1Client Selection in Federated Learning Non-IID Federated Learning
Samovar, T´el´ecom SudParis, Institut Polytechnique de Paris, France · ´ETS Montr´eal · CEA-LIST
SceneSelect: Selective Learning for Trajectory Scene Classification and Expert Scheduling
Apr 27, 2026Xinrun Wang, Deshun Xia, Yuxi Sun +1Multi-Agent Trajectory Prediction Expert Routing
School of Computer Science, China University of Geosciences (Wuhan) · School of Information Engineering, Wuhan University of Technology · School of Mathematics and Statistics, Wuhan University of Technology
CTP-FL: Common-Trajectory Gradient Prediction for Federated Learning
Sep 28, 2026Junkang LiuFederated Learning
Tianjin University

May 20, 2026, cs.LGJ/K move · Enter open · S save

Choose Wisely and Privately: Proactive Client Selection for Fair and Efficient Federated Learning

Adda Akram Bendoukha, Heber Hwang Arcolezi, Nesrine Kaaniche, Aymen Boudguiga

Samovar, T´el´ecom SudParis, Institut Polytechnique de Paris, France · ´ETS Montr´eal · CEA-LIST

Client Selection in Federated Learning Non-IID Federated Learning Federated Learning

Open paper AskPDF arXiv

Federated Learning enables collaborative model training across decentralized data sources without data transfer. Averaging-based FL is limited by the presence of non-IID data, which negatively impacts convergence speed and final model accuracy. Conventional alternatives suffer from significant inefficiency. Clients with noisy or highly heterogeneous data contribute expensive gradient computations that are either discarded or heavily down-weighted before aggregation. These reactive approaches waste computational resources, require more communication rounds and result in unnecessary privacy exposure. In this paper, we propose a proactive client selection framework that aims to find an optimal federation of clients whose combined data match utility and fairness requirements before training begins. Our method relies on mutual information computed from differentially private contingency tables to quantify the relevance of cross-feature correlations in the union dataset. We introduce a Potential Federation Loss (PFL) over the set of fixed-size federations, which balances two objectives. Maximizing collective data utility while ensuring fair cross-features correlations to prevent group unfairness. Client selection is expressed as an optimal subset search problem over the PFL objective, which we solve using simulated annealing under strong differential privacy guarantees for clients' local statistics. Experimental results on four benchmarks show faster, fairer, and more accurate models trained on optimally found federations, compared to uniform sampling, even when state-of-the-art adaptive aggregation or sampling strategies are employed.

Abstract

Explore similar work

Choose Wisely and Privately: Proactive Client Selection for Fair and Efficient Federated Learning

SceneSelect: Selective Learning for Trajectory Scene Classification and Expert Scheduling

CTP-FL: Common-Trajectory Gradient Prediction for Federated Learning