Learning Demographic-Conditioned Mobility Trajectories with Aggregate Supervision

Jessie Z. Li; Zhiqing Hong; Toru Shirakawa; Serina Chang

cs.LGMar 3, 2026

Learning Demographic-Conditioned Mobility Trajectories with Aggregate Supervision

Authors: Jessie Z. Li, Zhiqing Hong, Toru Shirakawa, Serina Chang

Organizations: University of California, Berkeley · Hong Kong University of Science and Technology (Guangzhou)

Abstract

Human mobility trajectories are widely studied in public health and social science, where different demographic groups exhibit significantly different mobility patterns. However, existing trajectory generation models rarely capture this heterogeneity because most trajectory datasets lack demographic labels. To address this gap in data, we propose ATLAS, a weakly supervised approach for demographic-conditioned trajectory generation using only (i) individual trajectories without demographic labels, (ii) region-level aggregated mobility features, and (iii) region-level demographic compositions from census data. ATLAS trains a trajectory generator and fine-tunes it so that simulated mobility matches observed regional aggregates while conditioning on demographics. Experiments on real trajectory data with demographic labels show that ATLAS substantially improves demographic realism over baselines (JSD $\downarrow$ 12%--69%) and closes much of the gap to strongly supervised training. We further develop theoretical analyses for when and why ATLAS works, identifying key factors including demographic diversity across regions and the informativeness of the aggregate feature, paired with experiments demonstrating the practical implications of our theory. We release our code at https://github.com/schang-lab/ATLAS.

Figures & tables

Figure 1. Overview of ATLAS . Phase 1: Train a generative model on trajectories without demographic labels. Phase 2: Fine-tune with demographic conditioning by sampling groups from the region’s demographic composition $p(\cdot\mid g)$ and optimizing to match region’s observed aggregate features $\nu_{\star}(g)$ .

Figure 2. Effect of demographic diversity on ATLAS performance (RQ1). JSD (lower is better) across four metrics on Virginia (top) and California (bottom). Bars indicate average JSD over 8 demographic groups; errors indicate standard deviation over groups. ATLAS substantially improves over the baseline under well-conditioned regional partitions, often approaching strongly supervised performance, and degrades gracefully as partitions become ill-conditioned.

State	Feature Map $\phi$	Spatial	Travel Dist.	Trip	POI Freq.
VA	Baseline	$0.105\pm\scriptstyle 0.019$	$0.382\pm\scriptstyle 0.045$	$0.157\pm\scriptstyle 0.030$	$0.304\pm\scriptstyle 0.016$
	Strong	$0.059\pm\scriptstyle 0.013$	$0.158\pm\scriptstyle 0.043$	$0.122\pm\scriptstyle 0.023$	$0.249\pm\scriptstyle 0.031$
	POI-Histogram	$0.080\pm\scriptstyle 0.008$	$0.124\pm\scriptstyle 0.027$	$0.105\pm\scriptstyle 0.019$	$0.211\pm\scriptstyle 0.023$
	Cate-Trans	$0.109\pm\scriptstyle 0.028$	$0.227\pm\scriptstyle 0.039$	$0.129\pm\scriptstyle 0.028$	$0.257\pm\scriptstyle 0.027$
	Cate	$0.103\pm\scriptstyle 0.026$	$0.344\pm\scriptstyle 0.119$	$0.149\pm\scriptstyle 0.036$	$0.278\pm\scriptstyle 0.040$
CA	Baseline	$0.102\pm\scriptstyle 0.007$	$0.565\pm\scriptstyle 0.038$	$0.106\pm\scriptstyle 0.011$	$0.224\pm\scriptstyle 0.013$

Table 1. Effect of feature choice on ATLAS performance (RQ2). Average JSD

\downarrow

across 8 demographic groups (

\pm

std across groups), for Virginia (VA) and California (CA).

Accuracy $\uparrow$
	Setup	$<$ 30, M	$<$ 30, F	30–40, M	30–40, F	40–50, M	40–50, F	$>$ 50, M	$>$ 50, F	Avg
VA	Real	0.587	0.638	0.549	0.568	0.570	0.523	0.565	0.521	0.565
	Baseline	0.480	0.473	0.520	0.555	0.474	0.440	0.417	0.437	0.475
	ATLAS	0.578	0.627	0.539	0.560	0.548	0.511	0.535	0.508	0.551
CA	Real	0.610	0.657	0.604	0.618	0.620	0.554	0.596	0.581	0.605
	Baseline	0.599	0.625	0.596	0.562	0.587	0.539	0.530	0.509	0.581

Table 2. Next-POI prediction (RQ3). Accuracy

\uparrow

and GeoError

\downarrow

(km) per group for Virginia (VA) and California (CA).

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition
$p_{g}\in\Delta^{K-1}$	region- $g$ demographic composition (row of $P$ )
$P\in\mathbb{R}^{G\times K}$	demographic composition matrix with rows $p_{g}^{\top}$
$\mu_{\star}(d)\in\mathbb{R}^{m}$	group- $d$ feature mean $\mathbb{E}_{P_{\star}(\cdot\mid d)}[\phi(X)]$
$\mu_{\theta}(d)\in\mathbb{R}^{m}$	model group- $d$ feature mean $\mathbb{E}_{P_{\theta}(\cdot\mid d)}[\phi(X)]$
$M_{\star}\in\mathbb{R}^{m\times K}$	columns are $\mu_{\star}(d)$
$M_{\theta}\in\mathbb{R}^{m\times K}$	columns are $\mu_{\theta}(d)$

Appendix

Table A1 . Key Notation.

Figure B1 . Pairwise JSD between demographic groups. Heatmaps showing pairwise Jensen–Shannon divergence (JSD) of POI visit distributions between $K{=}8$ age $\times$ gender groups for Virginia (left) and California (right). JSD values range from approximately 0.40 to 0.54, with warmer colors indicating larger divergence.

Shared Diffusion Settings
Diffusion steps	1000	Noise schedule	Cosine
Optimizer	Adam	Prediction type	$x_{0}$
BART		DiT
Encoder layers	4	DiT depth	8–12
Decoder layers	2	Hidden size	128–512
Model dimension	256	Attention heads	4

Appendix

Table B1 . Hyperparameters of our trajectory generation model.

Figure B2 . Virginia: average JSD during finetuning. Validation loss (TV) and per-metric JSD (averaged across demographic groups) over training steps.

Figure B3 . Virginia: per-group JSD during finetuning. Per-metric JSD broken down by demographic group over training steps.

Figure B4 . California: average JSD during finetuning. Validation loss (TV) and per-metric test JSD (averaged across demographic groups) over training steps.

Figure B5 . California: per-group JSD during finetuning. Per-metric JSD broken down by demographic group over training steps.

	Spatial JSD $\downarrow$		Travel Dist. JSD $\downarrow$		Trip JSD $\downarrow$		POI Freq. JSD $\downarrow$
Setup	Value	Gap Closed (%)	Value	Gap Closed (%)	Value	Gap Closed (%)	Value	Gap Closed (%)
Baseline (no demo)	$0.105\pm\scriptstyle 0.019$	–	$0.382\pm\scriptstyle 0.045$	–	$0.157\pm\scriptstyle 0.030$	–	$0.304\pm\scriptstyle 0.016$	–
Strong (supervised)	$0.059\pm\scriptstyle 0.013$	100	$0.158\pm\scriptstyle 0.043$	100	$0.122\pm\scriptstyle 0.023$	100	$0.249\pm\scriptstyle 0.031$	100
Demogroups-JS	$0.080\pm\scriptstyle 0.008$	$54$	$0.124\pm\scriptstyle 0.027$	$115$	$0.105\pm\scriptstyle 0.019$	$149$	$0.211\pm\scriptstyle 0.023$	$169$
Demogroups-TV	$0.087\pm\scriptstyle 0.018$	$39$	$0.131\pm\scriptstyle 0.067$	$112$	$0.125\pm\scriptstyle 0.017$	$91$	$0.238\pm\scriptstyle 0.026$	$120$
FullRank-JS	$0.090\pm\scriptstyle 0.016$	$33$	$0.167\pm\scriptstyle 0.037$	$96$	$0.121\pm\scriptstyle 0.021$	$103$	$0.243\pm\scriptstyle 0.022$	$111$

Appendix

Table C1 . Effect of demographic diversity (RQ1), Virginia. Average metrics across demographic groups for all partition structures (mean

\pm

std over 8 groups). Lower is better for JSD and higher is better for Gap Closed. Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.131\pm\scriptstyle 0.002$	$0.089\pm\scriptstyle 0.002$	$0.119\pm\scriptstyle 0.009$	$0.077\pm\scriptstyle 0.008$	$0.105\pm\scriptstyle 0.015$	$0.099\pm\scriptstyle 0.011$	$0.093\pm\scriptstyle 0.008$	$0.123\pm\scriptstyle 0.001$	$0.105\pm\scriptstyle 0.019$
Strong	$0.075\pm\scriptstyle 0.006$	$0.051\pm\scriptstyle 0.007$	$0.045\pm\scriptstyle 0.001$	$0.048\pm\scriptstyle 0.000$	$0.058\pm\scriptstyle 0.004$	$0.068\pm\scriptstyle 0.002$	$0.051\pm\scriptstyle 0.000$	$0.079\pm\scriptstyle 0.006$	$0.059\pm\scriptstyle 0.013$
Demogroups-JS	$0.072\pm\scriptstyle 0.018$	$0.060\pm\scriptstyle 0.008$	$0.092\pm\scriptstyle 0.011$	$0.057\pm\scriptstyle 0.029$	$0.086\pm\scriptstyle 0.010$	$0.093\pm\scriptstyle 0.008$	$0.091\pm\scriptstyle 0.008$	$0.089\pm\scriptstyle 0.037$	$0.080\pm\scriptstyle 0.008$
Demogroups-TV	$0.075\pm\scriptstyle 0.025$	$0.072\pm\scriptstyle 0.030$	$0.110\pm\scriptstyle 0.045$	$0.066\pm\scriptstyle 0.022$	$0.101\pm\scriptstyle 0.051$	$0.095\pm\scriptstyle 0.018$	$0.106\pm\scriptstyle 0.049$	$0.071\pm\scriptstyle 0.012$	$0.087\pm\scriptstyle 0.018$
FullRank-JS	$0.105\pm\scriptstyle 0.028$	$0.076\pm\scriptstyle 0.033$	$0.097\pm\scriptstyle 0.011$	$0.058\pm\scriptstyle 0.006$	$0.098\pm\scriptstyle 0.045$	$0.099\pm\scriptstyle 0.014$	$0.102\pm\scriptstyle 0.009$	$0.084\pm\scriptstyle 0.001$	$0.090\pm\scriptstyle 0.016$
FullRank-TV	$0.074\pm\scriptstyle 0.005$	$0.054\pm\scriptstyle 0.006$	$0.102\pm\scriptstyle 0.017$	$0.057\pm\scriptstyle 0.009$	$0.108\pm\scriptstyle 0.035$	$0.099\pm\scriptstyle 0.004$	$0.164\pm\scriptstyle 0.042$	$0.101\pm\scriptstyle 0.013$	$0.095\pm\scriptstyle 0.035$

Appendix

Table C2 . Effect of demographic diversity (RQ1), Virginia - spatial JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.398\pm\scriptstyle 0.038$	$0.353\pm\scriptstyle 0.075$	$0.455\pm\scriptstyle 0.053$	$0.324\pm\scriptstyle 0.074$	$0.366\pm\scriptstyle 0.070$	$0.350\pm\scriptstyle 0.081$	$0.387\pm\scriptstyle 0.059$	$0.425\pm\scriptstyle 0.045$	$0.382\pm\scriptstyle 0.045$
Strong	$0.125\pm\scriptstyle 0.002$	$0.178\pm\scriptstyle 0.003$	$0.140\pm\scriptstyle 0.003$	$0.134\pm\scriptstyle 0.002$	$0.147\pm\scriptstyle 0.003$	$0.164\pm\scriptstyle 0.008$	$0.121\pm\scriptstyle 0.004$	$0.253\pm\scriptstyle 0.020$	$0.158\pm\scriptstyle 0.043$
Demogroups-JS	$0.127\pm\scriptstyle 0.078$	$0.121\pm\scriptstyle 0.072$	$0.146\pm\scriptstyle 0.066$	$0.151\pm\scriptstyle 0.013$	$0.160\pm\scriptstyle 0.096$	$0.084\pm\scriptstyle 0.029$	$0.106\pm\scriptstyle 0.070$	$0.097\pm\scriptstyle 0.005$	$0.124\pm\scriptstyle 0.027$
Demogroups-TV	$0.079\pm\scriptstyle 0.043$	$0.099\pm\scriptstyle 0.040$	$0.081\pm\scriptstyle 0.054$	$0.282\pm\scriptstyle 0.045$	$0.149\pm\scriptstyle 0.015$	$0.090\pm\scriptstyle 0.046$	$0.155\pm\scriptstyle 0.087$	$0.116\pm\scriptstyle 0.062$	$0.131\pm\scriptstyle 0.067$
FullRank-JS	$0.180\pm\scriptstyle 0.117$	$0.169\pm\scriptstyle 0.096$	$0.105\pm\scriptstyle 0.058$	$0.144\pm\scriptstyle 0.023$	$0.234\pm\scriptstyle 0.138$	$0.173\pm\scriptstyle 0.009$	$0.178\pm\scriptstyle 0.135$	$0.150\pm\scriptstyle 0.028$	$0.167\pm\scriptstyle 0.037$
FullRank-TV	$0.075\pm\scriptstyle 0.003$	$0.114\pm\scriptstyle 0.021$	$0.150\pm\scriptstyle 0.009$	$0.146\pm\scriptstyle 0.054$	$0.210\pm\scriptstyle 0.106$	$0.162\pm\scriptstyle 0.084$	$0.313\pm\scriptstyle 0.110$	$0.219\pm\scriptstyle 0.061$	$0.174\pm\scriptstyle 0.073$

Appendix

Table C3 . Effect of demographic diversity (RQ1), Virginia - travel distance JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.192\pm\scriptstyle 0.006$	$0.122\pm\scriptstyle 0.005$	$0.177\pm\scriptstyle 0.010$	$0.119\pm\scriptstyle 0.005$	$0.165\pm\scriptstyle 0.009$	$0.170\pm\scriptstyle 0.006$	$0.128\pm\scriptstyle 0.018$	$0.186\pm\scriptstyle 0.014$	$0.157\pm\scriptstyle 0.030$
Strong	$0.146\pm\scriptstyle 0.000$	$0.108\pm\scriptstyle 0.016$	$0.113\pm\scriptstyle 0.002$	$0.092\pm\scriptstyle 0.004$	$0.127\pm\scriptstyle 0.007$	$0.154\pm\scriptstyle 0.000$	$0.098\pm\scriptstyle 0.001$	$0.140\pm\scriptstyle 0.020$	$0.122\pm\scriptstyle 0.023$
Demogroups-JS	$0.112\pm\scriptstyle 0.027$	$0.068\pm\scriptstyle 0.008$	$0.101\pm\scriptstyle 0.008$	$0.095\pm\scriptstyle 0.045$	$0.121\pm\scriptstyle 0.029$	$0.127\pm\scriptstyle 0.014$	$0.115\pm\scriptstyle 0.017$	$0.100\pm\scriptstyle 0.058$	$0.105\pm\scriptstyle 0.019$
Demogroups-TV	$0.133\pm\scriptstyle 0.031$	$0.092\pm\scriptstyle 0.030$	$0.138\pm\scriptstyle 0.050$	$0.105\pm\scriptstyle 0.032$	$0.142\pm\scriptstyle 0.088$	$0.131\pm\scriptstyle 0.033$	$0.131\pm\scriptstyle 0.061$	$0.127\pm\scriptstyle 0.052$	$0.125\pm\scriptstyle 0.017$
FullRank-JS	$0.153\pm\scriptstyle 0.040$	$0.098\pm\scriptstyle 0.024$	$0.113\pm\scriptstyle 0.013$	$0.096\pm\scriptstyle 0.014$	$0.136\pm\scriptstyle 0.052$	$0.141\pm\scriptstyle 0.028$	$0.122\pm\scriptstyle 0.029$	$0.108\pm\scriptstyle 0.014$	$0.121\pm\scriptstyle 0.021$
FullRank-TV	$0.118\pm\scriptstyle 0.006$	$0.082\pm\scriptstyle 0.022$	$0.112\pm\scriptstyle 0.008$	$0.071\pm\scriptstyle 0.014$	$0.144\pm\scriptstyle 0.054$	$0.139\pm\scriptstyle 0.026$	$0.207\pm\scriptstyle 0.034$	$0.135\pm\scriptstyle 0.013$	$0.126\pm\scriptstyle 0.042$

Appendix

Table C4 . Effect of demographic diversity (RQ1), Virginia - trip JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.307\pm\scriptstyle 0.005$	$0.291\pm\scriptstyle 0.010$	$0.292\pm\scriptstyle 0.007$	$0.291\pm\scriptstyle 0.018$	$0.308\pm\scriptstyle 0.022$	$0.318\pm\scriptstyle 0.018$	$0.290\pm\scriptstyle 0.000$	$0.332\pm\scriptstyle 0.002$	$0.304\pm\scriptstyle 0.016$
Strong	$0.252\pm\scriptstyle 0.005$	$0.243\pm\scriptstyle 0.001$	$0.189\pm\scriptstyle 0.001$	$0.242\pm\scriptstyle 0.001$	$0.262\pm\scriptstyle 0.006$	$0.279\pm\scriptstyle 0.001$	$0.232\pm\scriptstyle 0.003$	$0.291\pm\scriptstyle 0.001$	$0.249\pm\scriptstyle 0.031$
Demogroups-JS	$0.186\pm\scriptstyle 0.018$	$0.168\pm\scriptstyle 0.020$	$0.220\pm\scriptstyle 0.016$	$0.202\pm\scriptstyle 0.024$	$0.227\pm\scriptstyle 0.037$	$0.224\pm\scriptstyle 0.000$	$0.229\pm\scriptstyle 0.031$	$0.231\pm\scriptstyle 0.029$	$0.211\pm\scriptstyle 0.023$
Demogroups-TV	$0.211\pm\scriptstyle 0.049$	$0.204\pm\scriptstyle 0.077$	$0.223\pm\scriptstyle 0.028$	$0.261\pm\scriptstyle 0.057$	$0.262\pm\scriptstyle 0.059$	$0.245\pm\scriptstyle 0.028$	$0.273\pm\scriptstyle 0.080$	$0.222\pm\scriptstyle 0.012$	$0.238\pm\scriptstyle 0.026$
FullRank-JS	$0.262\pm\scriptstyle 0.041$	$0.207\pm\scriptstyle 0.065$	$0.229\pm\scriptstyle 0.010$	$0.228\pm\scriptstyle 0.009$	$0.277\pm\scriptstyle 0.087$	$0.250\pm\scriptstyle 0.028$	$0.255\pm\scriptstyle 0.042$	$0.235\pm\scriptstyle 0.010$	$0.243\pm\scriptstyle 0.022$
FullRank-TV	$0.215\pm\scriptstyle 0.014$	$0.169\pm\scriptstyle 0.013$	$0.246\pm\scriptstyle 0.029$	$0.212\pm\scriptstyle 0.029$	$0.318\pm\scriptstyle 0.055$	$0.274\pm\scriptstyle 0.016$	$0.378\pm\scriptstyle 0.042$	$0.255\pm\scriptstyle 0.034$	$0.258\pm\scriptstyle 0.066$

Appendix

Table C5 . Effect of demographic diversity (RQ1), Virginia - POI frequency JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

	Spatial JSD $\downarrow$		Travel Dist. JSD $\downarrow$		Trip JSD $\downarrow$		POI Freq. JSD $\downarrow$
Setup	Value	Gap Closed (%)	Value	Gap Closed (%)	Value	Gap Closed (%)	Value	Gap Closed (%)
Baseline	$0.102\pm\scriptstyle 0.007$	–	$0.565\pm\scriptstyle 0.038$	–	$0.106\pm\scriptstyle 0.011$	–	$0.224\pm\scriptstyle 0.013$	–
Strong	$0.070\pm\scriptstyle 0.017$	100	$0.253\pm\scriptstyle 0.046$	100	$0.090\pm\scriptstyle 0.013$	100	$0.180\pm\scriptstyle 0.026$	100
Demogroups-JS	$0.090\pm\scriptstyle 0.015$	$38$	$0.211\pm\scriptstyle 0.114$	$113$	$0.087\pm\scriptstyle 0.018$	$119$	$0.170\pm\scriptstyle 0.032$	$123$
Demogroups-TV	$0.087\pm\scriptstyle 0.013$	$47$	$0.210\pm\scriptstyle 0.080$	$114$	$0.100\pm\scriptstyle 0.025$	$37$	$0.173\pm\scriptstyle 0.042$	$116$
FullRank-JS	$0.089\pm\scriptstyle 0.015$	$41$	$0.177\pm\scriptstyle 0.068$	$124$	$0.081\pm\scriptstyle 0.008$	$156$	$0.161\pm\scriptstyle 0.014$	$143$

Appendix

Table C6 . Effect of demographic diversity (RQ1), California. Average metrics across demographic groups for all partition structures (mean

\pm

std over 8 groups). Lower is better for JSD and higher is better for Gap Closed. Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.096\pm\scriptstyle 0.003$	$0.095\pm\scriptstyle 0.011$	$0.101\pm\scriptstyle 0.003$	$0.096\pm\scriptstyle 0.007$	$0.110\pm\scriptstyle 0.000$	$0.104\pm\scriptstyle 0.001$	$0.115\pm\scriptstyle 0.003$	$0.103\pm\scriptstyle 0.008$	$0.102\pm\scriptstyle 0.007$
Strong	$0.056\pm\scriptstyle 0.002$	$0.052\pm\scriptstyle 0.000$	$0.067\pm\scriptstyle 0.001$	$0.060\pm\scriptstyle 0.002$	$0.105\pm\scriptstyle 0.004$	$0.072\pm\scriptstyle 0.001$	$0.079\pm\scriptstyle 0.002$	$0.069\pm\scriptstyle 0.007$	$0.070\pm\scriptstyle 0.017$
Demogroups-JS	$0.102\pm\scriptstyle 0.043$	$0.077\pm\scriptstyle 0.003$	$0.078\pm\scriptstyle 0.021$	$0.073\pm\scriptstyle 0.028$	$0.085\pm\scriptstyle 0.000$	$0.111\pm\scriptstyle 0.020$	$0.111\pm\scriptstyle 0.009$	$0.084\pm\scriptstyle 0.001$	$0.090\pm\scriptstyle 0.015$
Demogroups-TV	$0.075\pm\scriptstyle 0.002$	$0.072\pm\scriptstyle 0.001$	$0.090\pm\scriptstyle 0.005$	$0.073\pm\scriptstyle 0.017$	$0.094\pm\scriptstyle 0.007$	$0.108\pm\scriptstyle 0.009$	$0.098\pm\scriptstyle 0.019$	$0.086\pm\scriptstyle 0.002$	$0.087\pm\scriptstyle 0.013$
FullRank-JS	$0.085\pm\scriptstyle 0.019$	$0.085\pm\scriptstyle 0.030$	$0.123\pm\scriptstyle 0.038$	$0.080\pm\scriptstyle 0.036$	$0.083\pm\scriptstyle 0.021$	$0.084\pm\scriptstyle 0.000$	$0.098\pm\scriptstyle 0.016$	$0.073\pm\scriptstyle 0.007$	$0.089\pm\scriptstyle 0.015$
FullRank-TV	$0.060\pm\scriptstyle 0.003$	$0.073\pm\scriptstyle 0.000$	$0.085\pm\scriptstyle 0.000$	$0.066\pm\scriptstyle 0.008$	$0.109\pm\scriptstyle 0.000$	$0.111\pm\scriptstyle 0.001$	$0.095\pm\scriptstyle 0.005$	$0.098\pm\scriptstyle 0.001$	$0.087\pm\scriptstyle 0.019$

Appendix

Table C7 . Effect of demographic diversity (RQ1), California - spatial JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.502\pm\scriptstyle 0.008$	$0.583\pm\scriptstyle 0.015$	$0.549\pm\scriptstyle 0.003$	$0.568\pm\scriptstyle 0.010$	$0.595\pm\scriptstyle 0.008$	$0.523\pm\scriptstyle 0.012$	$0.614\pm\scriptstyle 0.012$	$0.583\pm\scriptstyle 0.034$	$0.565\pm\scriptstyle 0.038$
Strong	$0.207\pm\scriptstyle 0.020$	$0.232\pm\scriptstyle 0.015$	$0.227\pm\scriptstyle 0.006$	$0.285\pm\scriptstyle 0.027$	$0.342\pm\scriptstyle 0.047$	$0.205\pm\scriptstyle 0.032$	$0.279\pm\scriptstyle 0.008$	$0.248\pm\scriptstyle 0.027$	$0.253\pm\scriptstyle 0.046$
Demogroups-JS	$0.285\pm\scriptstyle 0.011$	$0.203\pm\scriptstyle 0.009$	$0.145\pm\scriptstyle 0.010$	$0.051\pm\scriptstyle 0.001$	$0.359\pm\scriptstyle 0.003$	$0.156\pm\scriptstyle 0.015$	$0.366\pm\scriptstyle 0.004$	$0.126\pm\scriptstyle 0.005$	$0.211\pm\scriptstyle 0.114$
Demogroups-TV	$0.238\pm\scriptstyle 0.128$	$0.135\pm\scriptstyle 0.058$	$0.185\pm\scriptstyle 0.101$	$0.201\pm\scriptstyle 0.059$	$0.143\pm\scriptstyle 0.021$	$0.182\pm\scriptstyle 0.013$	$0.389\pm\scriptstyle 0.199$	$0.208\pm\scriptstyle 0.074$	$0.210\pm\scriptstyle 0.080$
FullRank-JS	$0.206\pm\scriptstyle 0.012$	$0.107\pm\scriptstyle 0.030$	$0.178\pm\scriptstyle 0.066$	$0.086\pm\scriptstyle 0.055$	$0.116\pm\scriptstyle 0.017$	$0.261\pm\scriptstyle 0.128$	$0.261\pm\scriptstyle 0.058$	$0.201\pm\scriptstyle 0.012$	$0.177\pm\scriptstyle 0.068$
FullRank-TV	$0.119\pm\scriptstyle 0.014$	$0.185\pm\scriptstyle 0.008$	$0.191\pm\scriptstyle 0.015$	$0.106\pm\scriptstyle 0.001$	$0.396\pm\scriptstyle 0.030$	$0.157\pm\scriptstyle 0.020$	$0.254\pm\scriptstyle 0.003$	$0.197\pm\scriptstyle 0.023$	$0.201\pm\scriptstyle 0.092$

Appendix

Table C8 . Effect of demographic diversity (RQ1), California - travel distance JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.099\pm\scriptstyle 0.001$	$0.094\pm\scriptstyle 0.001$	$0.102\pm\scriptstyle 0.001$	$0.095\pm\scriptstyle 0.001$	$0.103\pm\scriptstyle 0.000$	$0.113\pm\scriptstyle 0.001$	$0.123\pm\scriptstyle 0.001$	$0.119\pm\scriptstyle 0.001$	$0.106\pm\scriptstyle 0.011$
Strong	$0.088\pm\scriptstyle 0.002$	$0.076\pm\scriptstyle 0.001$	$0.079\pm\scriptstyle 0.001$	$0.074\pm\scriptstyle 0.001$	$0.111\pm\scriptstyle 0.003$	$0.093\pm\scriptstyle 0.003$	$0.100\pm\scriptstyle 0.002$	$0.098\pm\scriptstyle 0.001$	$0.090\pm\scriptstyle 0.013$
Demogroups-JS	$0.112\pm\scriptstyle 0.002$	$0.076\pm\scriptstyle 0.001$	$0.071\pm\scriptstyle 0.000$	$0.063\pm\scriptstyle 0.001$	$0.089\pm\scriptstyle 0.002$	$0.097\pm\scriptstyle 0.002$	$0.111\pm\scriptstyle 0.000$	$0.081\pm\scriptstyle 0.009$	$0.087\pm\scriptstyle 0.018$
Demogroups-TV	$0.079\pm\scriptstyle 0.010$	$0.060\pm\scriptstyle 0.010$	$0.132\pm\scriptstyle 0.023$	$0.109\pm\scriptstyle 0.022$	$0.085\pm\scriptstyle 0.024$	$0.127\pm\scriptstyle 0.049$	$0.092\pm\scriptstyle 0.013$	$0.114\pm\scriptstyle 0.006$	$0.100\pm\scriptstyle 0.025$
FullRank-JS	$0.078\pm\scriptstyle 0.011$	$0.089\pm\scriptstyle 0.017$	$0.078\pm\scriptstyle 0.003$	$0.077\pm\scriptstyle 0.025$	$0.068\pm\scriptstyle 0.009$	$0.092\pm\scriptstyle 0.011$	$0.078\pm\scriptstyle 0.006$	$0.091\pm\scriptstyle 0.005$	$0.081\pm\scriptstyle 0.008$
FullRank-TV	$0.088\pm\scriptstyle 0.015$	$0.055\pm\scriptstyle 0.001$	$0.109\pm\scriptstyle 0.034$	$0.058\pm\scriptstyle 0.002$	$0.128\pm\scriptstyle 0.051$	$0.094\pm\scriptstyle 0.001$	$0.127\pm\scriptstyle 0.050$	$0.088\pm\scriptstyle 0.006$	$0.094\pm\scriptstyle 0.027$

Appendix

Table C9 . Effect of demographic diversity (RQ1), California - trip JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.216\pm\scriptstyle 0.003$	$0.198\pm\scriptstyle 0.008$	$0.222\pm\scriptstyle 0.005$	$0.218\pm\scriptstyle 0.013$	$0.232\pm\scriptstyle 0.012$	$0.233\pm\scriptstyle 0.011$	$0.236\pm\scriptstyle 0.010$	$0.234\pm\scriptstyle 0.012$	$0.224\pm\scriptstyle 0.013$
Strong	$0.160\pm\scriptstyle 0.001$	$0.135\pm\scriptstyle 0.000$	$0.172\pm\scriptstyle 0.000$	$0.168\pm\scriptstyle 0.002$	$0.216\pm\scriptstyle 0.002$	$0.194\pm\scriptstyle 0.003$	$0.202\pm\scriptstyle 0.001$	$0.190\pm\scriptstyle 0.006$	$0.180\pm\scriptstyle 0.026$
Demogroups-JS	$0.211\pm\scriptstyle 0.004$	$0.130\pm\scriptstyle 0.002$	$0.159\pm\scriptstyle 0.001$	$0.134\pm\scriptstyle 0.000$	$0.177\pm\scriptstyle 0.001$	$0.166\pm\scriptstyle 0.001$	$0.217\pm\scriptstyle 0.000$	$0.164\pm\scriptstyle 0.000$	$0.170\pm\scriptstyle 0.032$
Demogroups-TV	$0.132\pm\scriptstyle 0.001$	$0.106\pm\scriptstyle 0.016$	$0.223\pm\scriptstyle 0.011$	$0.185\pm\scriptstyle 0.014$	$0.154\pm\scriptstyle 0.006$	$0.225\pm\scriptstyle 0.008$	$0.175\pm\scriptstyle 0.006$	$0.187\pm\scriptstyle 0.016$	$0.173\pm\scriptstyle 0.042$
FullRank-JS	$0.143\pm\scriptstyle 0.015$	$0.161\pm\scriptstyle 0.040$	$0.157\pm\scriptstyle 0.013$	$0.155\pm\scriptstyle 0.021$	$0.148\pm\scriptstyle 0.007$	$0.184\pm\scriptstyle 0.021$	$0.177\pm\scriptstyle 0.007$	$0.164\pm\scriptstyle 0.014$	$0.161\pm\scriptstyle 0.014$
FullRank-TV	$0.160\pm\scriptstyle 0.010$	$0.157\pm\scriptstyle 0.051$	$0.184\pm\scriptstyle 0.017$	$0.172\pm\scriptstyle 0.026$	$0.209\pm\scriptstyle 0.027$	$0.210\pm\scriptstyle 0.009$	$0.199\pm\scriptstyle 0.026$	$0.184\pm\scriptstyle 0.017$	$0.184\pm\scriptstyle 0.021$

Appendix

Table C10 . Effect of demographic diversity (RQ1), California - POI frequency JSD per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups (matching Table C1 ). Bold indicates best ATLAS variant within each partition type.

Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Spatial JSD
Baseline	$0.131\pm\scriptstyle 0.002$	$0.089\pm\scriptstyle 0.002$	$0.119\pm\scriptstyle 0.009$	$0.077\pm\scriptstyle 0.008$	$0.105\pm\scriptstyle 0.015$	$0.099\pm\scriptstyle 0.011$	$0.093\pm\scriptstyle 0.008$	$0.123\pm\scriptstyle 0.001$	$0.105$
Strong	$0.075\pm\scriptstyle 0.006$	$0.051\pm\scriptstyle 0.007$	$0.045\pm\scriptstyle 0.001$	$0.048\pm\scriptstyle 0.000$	$0.058\pm\scriptstyle 0.004$	$0.068\pm\scriptstyle 0.002$	$0.051\pm\scriptstyle 0.000$	$0.079\pm\scriptstyle 0.006$	$0.059$
POI-Histogram-JS	$0.072\pm\scriptstyle 0.018$	$0.060\pm\scriptstyle 0.008$	$0.092\pm\scriptstyle 0.011$	$0.057\pm\scriptstyle 0.029$	$0.086\pm\scriptstyle 0.010$	$0.093\pm\scriptstyle 0.008$	$0.091\pm\scriptstyle 0.008$	$0.089\pm\scriptstyle 0.037$	$0.080$
POI-Histogram-TV	$0.075\pm\scriptstyle 0.025$	$0.072\pm\scriptstyle 0.030$	$0.110\pm\scriptstyle 0.045$	$0.066\pm\scriptstyle 0.022$	$0.101\pm\scriptstyle 0.051$	$0.095\pm\scriptstyle 0.018$	$0.106\pm\scriptstyle 0.049$	$0.071\pm\scriptstyle 0.012$	0.087
Cate-Trans-JS	$0.158\pm\scriptstyle 0.094$	$0.094\pm\scriptstyle 0.038$	$0.138\pm\scriptstyle 0.033$	$0.088\pm\scriptstyle 0.043$	$0.078\pm\scriptstyle 0.008$	$0.106\pm\scriptstyle 0.014$	$0.131\pm\scriptstyle 0.100$	$0.077\pm\scriptstyle 0.006$	0.109

Appendix

Table C11 . Effect of feature choice (RQ2), Virginia. JSDs across trajectory statistics per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups.

Spatial JSD
Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Baseline	$0.096\pm\scriptstyle 0.003$	$0.095\pm\scriptstyle 0.011$	$0.101\pm\scriptstyle 0.003$	$0.096\pm\scriptstyle 0.007$	$0.110\pm\scriptstyle 0.000$	$0.104\pm\scriptstyle 0.001$	$0.115\pm\scriptstyle 0.003$	$0.103\pm\scriptstyle 0.008$	$0.102$
Strong	$0.056\pm\scriptstyle 0.002$	$0.052\pm\scriptstyle 0.000$	$0.067\pm\scriptstyle 0.001$	$0.060\pm\scriptstyle 0.002$	$0.105\pm\scriptstyle 0.004$	$0.072\pm\scriptstyle 0.001$	$0.079\pm\scriptstyle 0.002$	$0.069\pm\scriptstyle 0.007$	$0.070$
POI-Histogram-JS	$0.102\pm\scriptstyle 0.043$	$0.077\pm\scriptstyle 0.003$	$0.078\pm\scriptstyle 0.021$	$0.073\pm\scriptstyle 0.028$	$0.085\pm\scriptstyle 0.000$	$0.111\pm\scriptstyle 0.020$	$0.111\pm\scriptstyle 0.009$	$0.084\pm\scriptstyle 0.001$	$0.090$
POI-Histogram-TV	$0.075\pm\scriptstyle 0.002$	$0.072\pm\scriptstyle 0.001$	$0.090\pm\scriptstyle 0.005$	$0.073\pm\scriptstyle 0.017$	$0.094\pm\scriptstyle 0.007$	$0.108\pm\scriptstyle 0.009$	$0.098\pm\scriptstyle 0.019$	$0.086\pm\scriptstyle 0.002$	$0.087$
Cate-Trans-JS	$0.094\pm\scriptstyle 0.016$	$0.091\pm\scriptstyle 0.038$	$0.094\pm\scriptstyle 0.017$	$0.120\pm\scriptstyle 0.067$	$0.117\pm\scriptstyle 0.015$	$0.118\pm\scriptstyle 0.040$	$0.162\pm\scriptstyle 0.066$	$0.155\pm\scriptstyle 0.020$	$0.119$

Appendix

Table C12 . Effect of feature choice (RQ2), California. JSDs across trajectory statistics per demographic group. Format: mean

\pm

std across 3 seeds. The Avg column reports mean

\pm

std across 8 demographic groups.

Accuracy
Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Real	0.5866	0.6377	0.5491	0.5677	0.5702	0.5225	0.5650	0.5212	0.5650
Baseline	0.4803	0.4729	0.5197	0.5553	0.4740	0.4404	0.4168	0.4371	0.4746
FullRank-JS	0.5777	0.6272	0.5386	0.5601	0.5482	0.5108	0.5348	0.5084	0.5507
FullRank-TV	0.5134	0.6292	0.4690	0.5637	0.5502	0.5030	0.4846	0.5087	0.5277
HR@10

Appendix

Table C13 . Next POI prediction performance by demographic group (Virginia). Metrics: Accuracy (Acc), Hit Rate at 10 (HR@10), Normalized Discounted Cumulative Gain at 10 (NDCG@10), and Geographic Error (GeoError in km). We report the mean across demographic groups in the last column.

Accuracy
Setup	<30, M	<30, F	30–40, M	30–40, F	40–50, M	40–50, F	>50, M	>50, F	Avg
Real	0.6100	0.6568	0.6043	0.6179	0.6202	0.5535	0.5955	0.5809	0.6049
Baseline	0.5990	0.6249	0.5964	0.5619	0.5867	0.5386	0.5301	0.5088	0.5808
FullRank-JS	0.5991	0.6453	0.5965	0.6126	0.6173	0.5395	0.5826	0.5598	0.5941
FullRank-TV	0.5995	0.6436	0.5966	0.6128	0.6168	0.5386	0.5820	0.5596	0.5937
HR@10

Appendix

Table C14 . Next POI prediction performance by demographic group (California). Metrics: Accuracy (Acc), Hit Rate at 10 (HR@10), Normalized Discounted Cumulative Gain at 10 (NDCG@10), and Geographic Error (GeoError in km). We report the mean across demographic groups in the last column.

Trajectory Generation

Trajectory generation creates plausible, goal-directed paths and motion sequences for people, vehicles, and robots.

See all 182 papers

Oct 20, 2026

Beyond Trajectory Statistics: Community-Consistent Human Mobility Generation via Evolving User-Location Graphs

Human mobility data are fundamental to Web-based location services and broader urban applications, and mobility generation provides synthetic trajectories when real data are costly or inaccessible. Although recent mobility generators have incorporated collective mobility patterns, they primarily focus on individual trajectories and aggregate population dynamics, leaving evolving community structures formed by user-location interactions insufficiently modeled. In this paper, we study community-consistent human mobility generation. To this end, we formulate mobility generation as the generation of dynamic user-location graphs to model community structures and propose MOTION, a Multi-agent framework for cOmmuniTy-consIstent human mObility geNeration. MOTION consists of user agents, location agents, and a global manager agent. MOTION first detects coherent communities from historical mobility using the Leiden algorithm and extracts community-conditioned interaction priors to guide the generation process. Guided by these priors, the user agents plan mobility, search destination candidates, and infer user movements from the individual view. In parallel, selectively activated location agents infer likely user arrivals from the community view. The manager coordinates both views, diagnoses interactions causing deviations, and attributes targeted reflection signals to revise user decisions, location attraction, and source-user selection. Experiments on two real-world datasets show that MOTION outperforms baselines by at least 4.05% across reported metrics of individual trajectories, population and flow distributions, and community and graph structures. Further analyses confirm its robustness, privacy preservation, and downstream utility. The code is available at https://anonymous.4open.science/r/MOTION-04C4.

Bo Liu, Tong Li, Jiapeng Zhang +3

National Supercomputing Center in Changsha Hunan University Changsha, China

Trajectory Generation Graph Generation

Ask PDF

Apr 7, 2026

ActivityEditor: Learning to Synthesize Physically Valid Human Mobility

Human mobility modeling is indispensable for diverse urban applications. However, existing data-driven methods often suffer from data scarcity, limiting their applicability in regions where historical trajectories are unavailable or restricted. To bridge this gap, we propose ActivityEditor, a novel dual-LLM-agent framework designed for zero-shot cross-regional trajectory generation. Our framework decomposes the complex synthesis task into two collaborative stages. Specifically, an intention-based agent, which leverages demographic-driven priors to generate structured human intentions and coarse activity chains to ensure high-level socio-semantic coherence. These outputs are then refined by editor agent to obtain mobility trajectories through iteratively revisions that enforces human mobility law. This capability is acquired through reinforcement learning with multiple rewards grounded in real-world physical constraints, allowing the agent to internalize mobility regularities and ensure high-fidelity trajectory generation. Extensive experiments demonstrate that ActivityEditor achieves superior zero-shot performance when transferred across diverse urban contexts. It maintains high statistical fidelity and physical validity, providing a robust and highly generalizable solution for mobility simulation in data-scarce scenarios. Our code is available at: https://anonymous.4open.science/r/ActivityEditor-066B.

Chenjie Yang, Yutian Jiang, Anqi Liang +3

Southwest Jiaotong University · HKUST (Guangzhou) · Shanghai Jiao Tong University +2

Trajectory Generation LLM Agent Trajectory Synthesis

Ask PDF

Jun 10, 2026

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

Human mobility data is important for transportation, urban planning, and epidemic control, but large-scale trajectory collection is often costly and privacy-constrained, motivating realistic synthetic trajectory generation. Existing LLM-based generators typically rely on either prompt engineering, which preserves zero-shot reasoning but lacks fine-grained spatiotemporal grounding, or trajectory-level fine-tuning, which improves statistical precision but incurs substantial computational cost and may weaken general reasoning. We propose TrajGenAgent, a semantic-aware hierarchical LLM-agent framework for human mobility trajectory generation without model fine-tuning. TrajGenAgent uses a two-stage orchestrator-worker design: an LLM first synthesizes an individual- and weekday-conditioned activity chain from historical evidence via in-context learning, and a deterministic workflow then grounds each activity into a complete visit using personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation. To evaluate realism beyond aggregate spatiotemporal statistics, we introduce an anomaly-detection-based evaluation framework using two complementary detectors to assess behavioral and semantic plausibility. Experiments on benchmark and large-scale simulation datasets show that TrajGenAgent improves spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism over representative neural and LLM-based baselines, while avoiding parameter updates.

Siyu Li, Toan Tran, Lingyi Zhao +2

Dept. of Computer Science, Emory University, Atlanta, GA, USA · Novateur Research Solutions, Ashburn, VA, USA

Trajectory Generation Human Behavior Simulation Synthetic Data Generation LLM Agent Trajectory Synthesis LLM Agent Planning

Ask PDF

Abstract

Trajectory Generation

Explore similar work

Beyond Trajectory Statistics: Community-Consistent Human Mobility Generation via Evolving User-Location Graphs

ActivityEditor: Learning to Synthesize Physically Valid Human Mobility

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation