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

Yushi Bai, Kang Liu, Ling Yin, Wei Wang, Yilin LyuarXiv 2026 · arXiv

TL;DR

Act-HiMotif is a new hierarchical generative framework that synthesizes realistic daily activity schedules based on individual profiles (like age and income). By introducing learned, reusable "activity-structure motifs," the model successfully navigates the highly imbalanced, long-tailed distribution of human behavior while ensuring every generated schedule fits perfectly within a strict 24-hour time budget.

What Problem Does This Paper Address?

Simulating how people spend their time—what they do, in what order, and for how long—is crucial for urban planning, travel demand modeling, and social science. However, creating these schedules is surprisingly difficult for two main reasons:

  1. The Long-Tail Problem: Human behavior is incredibly diverse. In the 2017 U.S. National Household Travel Survey (NHTS), over 90% of distinct activity sequences (e.g., Home-Work-Shop-Home) appear five times or fewer. Standard deep learning models struggle to learn these rare patterns.
  2. The 24-Hour Constraint: A valid schedule must sum to exactly 24 hours. Most generative models either ignore this global constraint or require messy "post-hoc" adjustments that ruin the realism of the schedule.

As shown in Figure 1, many rare structures actually share common "motifs" (like being work-centered or school-centered), and the same structure can have vastly different duration allocations.

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.

Methods

The authors propose Act-HiMotif, which breaks the generation process into a coarse-to-fine hierarchy:

  • Active-Day Gate: First, the model decides if the person stays home all day or goes out.
  • Vector-Quantized (VQ) Motif Learner: This is the "secret sauce." The model learns a discrete codebook of "motifs"—reusable templates of activity organization. This allows rare sequences to "borrow" statistical strength from more common ones that share the same motif.
  • Motif-Conditioned Decoder: Given a profile and a sampled motif, this module generates the specific sequence of activities.
  • Budget-Constrained Duration Allocator: Instead of predicting durations directly, this module predicts a sequence of "endpoints" throughout the day. By masking out impossible times (e.g., you can't end an activity before it starts or leave no time for remaining activities), it guarantees the final schedule sums to exactly 24 hours.

The full architecture is detailed in Figure 2.

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.

Datasets

The model was trained and evaluated using the 2017 U.S. National Household Travel Survey (NHTS).

  • Scale: 264,234 person-day schedules.
  • Features: 9 activity categories (Home, Work, School, etc.) and 48 half-hour time slots.
  • Evaluation: The authors tested both "in-distribution" (same regions) and "spatial generalization" (testing on entirely different cities/regions) to ensure the model wasn't just memorizing local data.

Key Insights and Ideas

  • Motifs as a Bridge: By using VQ-VAE style motifs, the model creates an intermediate representation. This allows it to recognize that a rare "Home-Work-Gym-Home" sequence is fundamentally similar to a common "Home-Work-Home" sequence, improving its ability to generate plausible rare behaviors.
  • Hierarchy Beats Flatness: Flat models that try to predict everything at once often fail to capture the nuances of different profile groups. Act-HiMotif's hierarchy allows it to better align specific chains with specific demographics (e.g., students and school-centered motifs).
  • Validity by Construction: The duration allocator's use of feasible-range masking ensures temporal validity without needing a separate optimization step, making it more efficient and robust than previous methods.

Key Results

Act-HiMotif significantly outperformed strong baselines like TrajGPT and ActVAE.

  • Structural Fidelity: It achieved an 80.2% improvement in Episode JSD (measuring sequence complexity) and a 5.2% boost in Transition CPC (measuring how well it captures activity transitions).
  • Temporal Precision: The gains were even more dramatic in duration allocation, with a 93% improvement in Duration JSD.
  • Handling the Long Tail: As shown in Figure 3, Act-HiMotif maintains high fidelity even for "Rare," "Singleton," and "Unseen" activity structures where other models' performance collapses.
Figure 3: Duration and transition fidelity by ground-truth structure support. Groups are defined by training frequency: Head (ftrain ≥100), Medium (10–99), Rare (2–9), Singleton (1), and Unseen (0).

Qualitatively, the model excels at capturing the joint distribution of when activities start and how long they last, as seen in Figure 5.

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

Limitations and Future Work

While Act-HiMotif is a major step forward, the authors note some areas for growth:

  • Spatial Details: The current model focuses on activity types and timing but does not yet generate specific geographic coordinates or travel modes (e.g., bus vs. car).
  • Static Profiles: The model uses fixed demographic attributes. Future iterations could incorporate more dynamic "need-based" triggers for activities.
  • Integration: A logical next step is integrating this generator into full-scale agent-based urban simulators to see how these realistic schedules impact city-wide traffic and energy patterns.

Summary updated

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.
BibTeX
@misc{bai2026act,
  author = {Yushi Bai and Kang Liu and Ling Yin and Wei Wang and Yilin Lyu},
  title = {{Act-HiMotif: Hierarchical Daily Activity Schedule Generation via Learned Activity-Structure Motifs}},
  year = {2026},
  eprint = {2611.06038},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.06038},
}
Figure