Act-HiMotif: Hierarchical Daily Activity Schedule Generation via Learned Activity-Structure Motifs

Yushi Bai; Kang Liu; Ling Yin; Wei Wang; Yilin Lyu

cs.LGNov 4, 2026

Act-HiMotif: Hierarchical Daily Activity Schedule Generation via Learned Activity-Structure Motifs

Authors: Yushi Bai, Kang Liu, Ling Yin, Wei Wang, Yilin Lyu

Organizations: Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China · School of Computer Science and Engineering, Macau University of Science and Technology, Macau, China · Shenzhen University of Advanced Technology, Shenzhen, China · University of Chinese Academy of Sciences, Beijing, China ·

Abstract

Generating daily activity schedules from individual attributes (e.g., age and income) is valuable for activity-based travel modeling and human behavior simulation. The task is challenging because exact activity structures follow a highly imbalanced, long-tailed distribution, while the same structure can admit multiple duration allocations under a fixed full-day budget. Direct generation must therefore learn many exact activity structures from highly uneven observations while modeling their possible temporal realizations. However, many exact structures share recurring activity organization. Based on this observation, we propose Act-HiMotif, a hierarchical motif-based generative model that introduces learned, reusable activity-structure motifs as an intermediate representation between individual profiles and specific activity structures. Act-HiMotif first distinguishes home-only days from active days, then samples a motif from a profile-conditioned prior to guide activity-structure generation, and finally applies a budget-constrained duration allocator to generate activity durations that sum exactly to the full-day horizon. This hierarchy models home-only routing, activity-structure generation, and duration allocation at their corresponding levels, while motifs allow related structures to share statistical support without losing their distinct sequences. Experiments on large-scale travel survey data show that Act-HiMotif outperforms strong baselines in both in-distribution and spatial generalization settings, demonstrating the effectiveness of hierarchical generation and reusable activity-structure motifs for producing realistic profile-conditioned daily activity schedules.

Figures & tables

Figure 1: Motivating properties of daily activity schedules: A. long-tailed activity structures with recurring patterns and B. multiple duration allocations for the same structure under the full-day time constraint.

Figure 2: Overview of Act-HiMotif: (a) hierarchical profile-conditioned schedule generation, (b) training-only learning of reusable activity-structure motifs, and (c) autoregressive budget-constrained duration allocation.

Method	In-distribution					Spatial generalization
Method	Epi. $\downarrow$	Dur. $\downarrow$	Start $\downarrow$	Slot $\downarrow$	CPC $\uparrow$	Epi. $\downarrow$	Dur. $\downarrow$	Start $\downarrow$	Slot $\downarrow$	CPC $\uparrow$
ConditionalRNN	0.3092	0.3646	0.4642	0.0311	0.4919	0.3468	0.3571	0.4496	0.0366	0.4479
GenerativeRNN	0.0100	0.0257	0.0334	0.0047	0.8900	0.0233	0.0385	0.0512	0.0063	0.8497
ActVAE	0.0314	0.0419	0.0325	0.0088	0.8591	0.0459	0.0571	0.0429	0.0071	0.8770
DeepActivity-ZS	0.2310	0.2817	0.3593	0.0269	0.5136	0.2200	0.2623	0.3722	0.0282	0.5092
ControlTraj	0.1492	0.1164	0.1016	0.0075	0.7184	0.1810	0.1220	0.1144	0.0078	0.7278

Table 1: Five-seed mean performance under in-distribution and spatial-generalization evaluation. Best and second-best results are bolded and underlined .

In-distribution
Variant	Epi. $\downarrow$	Dur. $\downarrow$	Start $\downarrow$	Slot $\downarrow$	CPC $\uparrow$
w/o Active-Day Gate ( $A$ )	0.0029	0.0016	0.0047	0.0002	0.9706
w/o Act.-Stru. Motifs ( $z$ )	0.0223	0.0051	0.0089	0.0021	0.8910
w/o Hie. Gen. ( $\mathbf{x}\!\rightarrow\!Y$ )	0.0029	0.0017	0.0069	0.0007	0.9067
Act-HiMotif	0.0014	0.0013	0.0045	0.0001	0.9749

Table 2: Five-seed mean ablation results for the hierarchical components of Act-HiMotif.

Figure 3: Duration and transition fidelity by ground-truth structure support. Groups are defined by training frequency: Head (

f_{\mathrm{train}}\geq 100

), Medium (

10

–

99

), Rare (

2

–

9

), Singleton (

1

), and Unseen (

0

Figure 4: Profile–activity-chain alignment under CBSA-disjoint spatial generalization. The top four active-day chains by training-set prevalence difference for the corresponding group.

Figure 5: Observed and generated joint start-time–duration distributions for four representative activities under CBSA-disjoint spatial generalization.

Dataset statistics
Statistic	Value
Person-day schedules	264,234
Activity categories	9
Temporal resolution	48 slots (30 min)
Profile fields	7
Distinct exact activity structures	20,425

Table 3: Processed dataset statistics and evaluation splits.

Figure 6: One-at-a-time sensitivity of Duration JSD and Transition CPC to the codebook size

M

, commitment weight

\beta

, and loss weights

\lambda_{A}

\lambda_{M}

\lambda_{S}

, and

\lambda_{D}

. Lines report five-seed means, and shaded regions indicate

\pm 1

standard deviation separately under each evaluation protocol.

Method	Target data	In-distribution					Spatial generalization
Method	Target data	Epi. $\downarrow$	Dur. $\downarrow$	Start $\downarrow$	Slot $\downarrow$	CPC $\uparrow$	Epi. $\downarrow$	Dur. $\downarrow$	Start $\downarrow$	Slot $\downarrow$	CPC $\uparrow$
No target-CBSA schedules
ConditionalRNN	0%	0.3092 $\pm$ 0.0074	0.3646 $\pm$ 0.0199	0.4642 $\pm$ 0.0179	0.0311 $\pm$ 0.0035	0.4919 $\pm$ 0.0256	0.3468 $\pm$ 0.0076	0.3571 $\pm$ 0.0314	0.4496 $\pm$ 0.0321	0.0366 $\pm$ 0.0046	0.4479 $\pm$ 0.0270
GenerativeRNN	0%	0.0100 $\pm$ 0.0021	0.0257 $\pm$ 0.0057	0.0334 $\pm$ 0.0053	0.0047 $\pm$ 0.0019	0.8900 $\pm$ 0.0207	0.0233 $\pm$ 0.0100	0.0385 $\pm$ 0.0094	0.0512 $\pm$ 0.0129	0.0063 $\pm$ 0.0033	0.8497 $\pm$ 0.0263
ActVAE	0%	0.0314 $\pm$ 0.0075	0.0419 $\pm$ 0.0097	0.0325 $\pm$ 0.0027	0.0088 $\pm$ 0.0022	0.8591 $\pm$ 0.0125	0.0459 $\pm$ 0.0059	0.0571 $\pm$ 0.0058	0.0429 $\pm$ 0.0046	0.0071 $\pm$ 0.0032	0.8770 $\pm$ 0.0288
DeepActivity-ZS	0%	0.2310 $\pm$ 0.0427	0.2817 $\pm$ 0.0436	0.3593 $\pm$ 0.0205	0.0269 $\pm$ 0.0045	0.5136 $\pm$ 0.0396	0.2200 $\pm$ 0.0440	0.2623 $\pm$ 0.0392	0.3722 $\pm$ 0.0104	0.0282 $\pm$ 0.0045	0.5092 $\pm$ 0.0692

Table 4: Complete five-seed results under in-distribution and spatial-generalization evaluation. Values are mean

\pm

standard deviation. Best and second-best results among methods using no target-CBSA schedules are shown in bold and underline , respectively.

In-distribution
Variant	Epi. $\downarrow$	Dur. $\downarrow$	Start $\downarrow$	Slot $\downarrow$	CPC $\uparrow$
w/o Active-Day Gate ( $A$ )	0.0029 $\pm$ 0.0006	0.0016 $\pm$ 0.0004	0.0047 $\pm$ 0.0004	0.00017 $\pm$ 0.00005	0.9706 $\pm$ 0.0038
w/o Act.-Stru. Motifs ( $z$ )	0.0223 $\pm$ 0.0048	0.0051 $\pm$ 0.0006	0.0089 $\pm$ 0.0008	0.00213 $\pm$ 0.00052	0.8910 $\pm$ 0.0098
w/o Hie. Gen. ( $\mathbf{x}\!\rightarrow\!Y$ )	0.0029 $\pm$ 0.0006	0.0017 $\pm$ 0.0003	0.0069 $\pm$ 0.0010	0.00070 $\pm$ 0.00011	0.9067 $\pm$ 0.0057
Act-HiMotif	0.0014 $\pm$ 0.0007	0.0013 $\pm$ 0.0001	0.0045 $\pm$ 0.0003	0.00011 $\pm$ 0.00006	0.9749 $\pm$ 0.0027

Table 5: Full five-seed ablation results reported as mean

\pm

standard deviation.

Figure 7: Frequency and cumulative coverage of the learned motifs. Bars report the mean individual coverage of the 20 most frequent motifs across five seeds, with error bars denoting one standard deviation. Curves and shaded regions show the corresponding mean cumulative coverage and across-seed variation. The in-distribution and spatial-generalization training sets contain 15,915 and 8,667 distinct exact activity structures, respectively, organized using a 256-entry motif codebook under each protocol.

Figure 8: Dominant activity structures of the four most frequent motifs for the fixed canonical seed 2026. Each row reports the motif-frequency rank, run-specific code index, individual and cumulative training coverage, and dominant exact activity structure.

Figure 9: Profile composition of the 16 most frequent motifs (frequency ranks 1–16) for the fixed canonical seed 2026 under each evaluation protocol. Rows are ordered by active-day training frequency and labeled with the motif index and dominant exact structure. For each motif, percentages sum to 100% separately across employment categories, day types, and age bands.

Figure 10: Profile–activity-chain alignment for student-status and age-band contrasts under CBSA-disjoint spatial generalization. Displayed active-day chains are selected using training-set prevalence differences.

Figure 11: Profile–activity-chain alignment for all analyzed profile contrasts under in-distribution evaluation. Displayed active-day chains are selected using training-set prevalence differences.

Figure 12: Profile-stratified Duration JSD and Transition CPC under CBSA-disjoint spatial generalization.

Figure 13: Profile-stratified Duration JSD and Transition CPC under in-distribution evaluation.

Figure 14: Observed and generated joint start-time–duration distributions for Home, Work, Shopping, and Recreation under in-distribution evaluation.

Figure 15: Observed and generated joint start-time–duration distributions for School, Medical, Escort, Meal, and Other under CBSA-disjoint spatial generalization.

Figure 16: Observed and generated joint start-time–duration distributions for School, Medical, Escort, Meal, and Other under in-distribution evaluation.

Figure 17: Row-normalized activity-transition distributions under CBSA-disjoint spatial generalization (top) and in-distribution evaluation (bottom). Rows denote current activities and columns denote next activities.

Generative Modeling

Generative modeling learns data distributions to create new samples, from images and text to molecules, simulations, and structured records.

See all 524 papers

Generating Realistic Individual Activity Schedules via Activity Location Allocation Based on Simulated Travel Times
Jun 23, 2026Tatsuya Mitomi, Yahya Gamal, Esra Suel +3Synthetic Data Generation
AI for Collective Intelligence Hub, University of Glasgow, Glasgow, UK · Fujitsu Limited, Kawasaki, Japan · Urban Analytics, Department of Geography, University of Zurich, Zurich, Switzerland +1
IBAD: Interpretable Behavioral Anomaly Detection on Human Mobility Data
Jun 14, 2026Bita Azarijoo, John Krumm, Cyrus ShahabiInterpretable Anomaly Detection Anomaly Detection
University of Southern California Los Angeles, California, USA
MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation
Oct 8, 2026Mengkun Gao, Zengqing Wu, Renhe Jiang +6Human Behavior Modeling Trajectory Generation
The University of Osaka · The University of Tokyo · The Hong Kong University of Science and Technology +1

Jun 23, 2026J/K move · Enter open · S save

Generating Realistic Individual Activity Schedules via Activity Location Allocation Based on Simulated Travel Times

Tatsuya Mitomi, Yahya Gamal, Esra Suel, Gary Polhill+2

AI for Collective Intelligence Hub, University of Glasgow, Glasgow, UK · Fujitsu Limited, Kawasaki, Japan · Urban Analytics, Department of Geography, University of Zurich, Zurich, Switzerland+1

Synthetic Data Generation

Open paper AskPDF arXiv

Individual level daily activity schedules are essential for a wide range of applications, including infectious disease control, urban transportation planning, and policy design. In practice, such schedules are typically generated by combining population data with travel survey data. These data sources are used because they are often publicly available, whereas observed individual activity schedules are difficult to obtain due to privacy concerns. However, because of the complexity of mobility modelling, it is difficult to generate realistic activity schedules that also preserve travel times consistent with those reported in travel surveys. To address this issue, we propose a framework for generating activity schedules that iteratively applies a dynamic programming method to allocate activity locations based on simulated travel times. Numerical experiments with dummy data show that the proposed method reduces the discrepancy between simulated travel times and those reported in travel surveys by 52.2% relative to the first iteration through iterative refinement.

Abstract

Generative Modeling

Explore similar work

Generating Realistic Individual Activity Schedules via Activity Location Allocation Based on Simulated Travel Times

IBAD: Interpretable Behavioral Anomaly Detection on Human Mobility Data

MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation