Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ActVAE: Modelling human activity schedules with a deep conditional generative approach

Fred Shone, Tim Hillel

TL;DR

ActVAE presents a conditional variational auto-encoder tailored to human activity scheduling, combining explicit density modelling with label-conditioned generation to capture both realistic distributions and scenario-dependent changes. The approach is evaluated against purely generative and purely conditional baselines using joint density estimation and MI analyses, showing superior performance in reproducing real schedule distributions while enabling counterfactual reasoning. The work demonstrates rapid training and generation, data-efficient learning, and practical applicability within transport demand frameworks, though it notes residual latent–label entanglement and some underestimation of conditional effects. Overall, ActVAE provides a robust, scalable tool for modeling diverse human schedules under varying demographic and policy conditions, with significant implications for demand modelling and scenario analysis.

Abstract

Modelling the complexity and diversity of human activity scheduling behaviour is inherently challenging. We demonstrate a deep conditional-generative machine learning approach for the modelling of realistic activity schedules depending on input labels such as an individual's age, employment status, or other information relevant to their scheduling. We combine (i) a structured latent generative approach, with (ii) a conditional approach, through a novel Conditional VAE architecture. This allows for the rapid generation of precise and realistic schedules for different input labels. We extensively evaluate model capabilities using a joint density estimation framework and several case studies. We additionally show that our approach has practical data and computational requirements, and can be deployed within new and existing demand modelling frameworks. We evaluate the importance of generative capability more generally, by comparing our combined approach to (i) a purely generative model without conditionality, and (ii) a purely conditional model which outputs the most likely schedule given the input labels. This comparison highlights the usefulness of explicitly modelling the randomness of complex and diverse human behaviours using deep generative approaches.

ActVAE: Modelling human activity schedules with a deep conditional generative approach

TL;DR

ActVAE presents a conditional variational auto-encoder tailored to human activity scheduling, combining explicit density modelling with label-conditioned generation to capture both realistic distributions and scenario-dependent changes. The approach is evaluated against purely generative and purely conditional baselines using joint density estimation and MI analyses, showing superior performance in reproducing real schedule distributions while enabling counterfactual reasoning. The work demonstrates rapid training and generation, data-efficient learning, and practical applicability within transport demand frameworks, though it notes residual latent–label entanglement and some underestimation of conditional effects. Overall, ActVAE provides a robust, scalable tool for modeling diverse human schedules under varying demographic and policy conditions, with significant implications for demand modelling and scenario analysis.

Abstract

Modelling the complexity and diversity of human activity scheduling behaviour is inherently challenging. We demonstrate a deep conditional-generative machine learning approach for the modelling of realistic activity schedules depending on input labels such as an individual's age, employment status, or other information relevant to their scheduling. We combine (i) a structured latent generative approach, with (ii) a conditional approach, through a novel Conditional VAE architecture. This allows for the rapid generation of precise and realistic schedules for different input labels. We extensively evaluate model capabilities using a joint density estimation framework and several case studies. We additionally show that our approach has practical data and computational requirements, and can be deployed within new and existing demand modelling frameworks. We evaluate the importance of generative capability more generally, by comparing our combined approach to (i) a purely generative model without conditionality, and (ii) a purely conditional model which outputs the most likely schedule given the input labels. This comparison highlights the usefulness of explicitly modelling the randomness of complex and diverse human behaviours using deep generative approaches.

Paper Structure

This paper contains 72 sections, 21 equations, 13 figures, 29 tables.

Table of Contents

  1. Introduction
  2. Literature review
  3. Activity scheduling model taxonomy
  4. Theory versus data-driven
  5. Compositional versus joint
  6. Conditional versus non-conditional
  7. Implicit versus explicit generative
  8. Evaluation of the literature
  9. Traditional compositional approaches
  10. Optimisation approaches
  11. Machine learning approaches
  12. Deep generative modelling approaches
  13. Motivations for our approach to activity scheduling
  14. Generative versus discriminative problems
  15. Critique of compositional and autoregressive approaches
  16. ...and 57 more sections

Figures (13)

  • Figure 1: Graphical models of generative (green) and inference (dashed) processes. (a) VAE, (b) CVAE with independent latent space, and (c) CVAE with dependent latent space, i.e. with latent space entangled with labels.
  • Figure 2: Model generative and conditional processed (a) ConditionalRNN (conditional), (b) GenerativeRNN (explicit generative), and (c) ActVAE (conditional-generative).
  • Figure 3: Example UK NTS Schedules
  • Figure 4: Summary of model encoder and decoder architectures and loss functions (a) ConditionalRNN (conditional), (b) GenerativeRNN (explicit generative), and (c) ActVAE (conditional-generative).
  • Figure 5: ActVAE architecture
  • ...and 8 more figures