Multitask learning deep neural networks to combine revealed and stated preference data

Shenhao Wang, Qingyi Wang, Jinhua ZhaoJournal of Choice Modelling 2020 · DOI

TL;DR

This paper introduces a novel Multitask Learning Deep Neural Network (MTLDNN) framework to effectively combine revealed preference (RP) and stated preference (SP) data for travel demand analysis. The MTLDNN framework significantly outperforms traditional methods like Nested Logit (NL) by achieving higher prediction accuracy and offering greater interpretability, demonstrating the power of deep learning for complex behavioral modeling.

What Problem Does This Paper Address?

Analyzing travel behavior often requires combining two types of data: revealed preference (RP) data, which reflects actual choices but can suffer from limited attribute coverage and high correlations, and stated preference (SP) data, which allows for testing new scenarios but is prone to biases from survey formats and hypothetical situations. Traditional methods like Nested Logit (NL) attempt to combine these data sources but rely heavily on handcrafted feature engineering and rigid assumptions. This paper addresses the challenge of creating a more flexible and powerful framework that can automatically learn features and incorporate soft constraints to better integrate RP and SP data for more accurate and insightful travel demand analysis.

Methods

The core of this paper is the proposed Multitask Learning Deep Neural Network (MTLDNN) framework. Unlike traditional methods that require manual feature engineering, MTLDNNs leverage deep neural networks to automatically learn relevant features from the data. The architecture consists of shared layers that capture commonalities across tasks (RP and SP) and task-specific layers that handle unique aspects of each. This "multitask" approach allows the model to learn from multiple related datasets simultaneously, improving generalization.

Key methodological innovations include:

  • Automatic Feature Learning: DNNs inherently learn feature representations, eliminating the need for manual feature engineering common in NL models.
  • Soft Constraints: Instead of hard constraints (e.g., forcing parameters to be equal), MTLDNNs use architectural design and regularization techniques to flexibly model the relationships between RP and SP. This includes controlling the similarity between task-specific layers.
  • Flexible Architecture: The MTLDNN architecture allows for varying numbers of shared and task-specific layers, offering a more adaptable model structure than the fixed architecture of NL models.

Datasets

The study utilizes a dataset collected in Singapore for analyzing travel demand, specifically focusing on the adoption of autonomous vehicles (AV). The survey included:

  • Revealed Preference (RP) data: Covering four existing travel modes (walking, public transit, driving, ride-sharing).
  • Stated Preference (SP) data: Including the four RP modes plus an additional mode, on-demand AV.
  • Socioeconomic variables: Collected to understand demographic influences.

In total, 1,592 RP choice answers and 8,418 SP choice answers were gathered. The data was split into training and testing sets with a 4:1 ratio.

Key Insights and Ideas

  • MTLDNNs are more generic and flexible than NL: The ability to automatically learn features and use soft constraints makes MTLDNNs a more powerful and adaptable tool for combining RP and SP data compared to the rigid, handcrafted nature of NL models.
  • Soft constraints are crucial for performance: The significant performance gains of MTLDNNs over standard DNNs are primarily attributed to the MTLDNN-specific soft constraints (architectural design and regularization), rather than just the automatic feature learning capacity of DNNs.
  • AV adoption is highly sensitive to cost and time: The interpretability of MTLDNNs reveals that AV adoption is strongly influenced by cost and travel time variables, with driving being a major substitute. Socioeconomic factors like age and income play a less significant role.
  • Deep learning doesn't automatically guarantee improvement: Simply applying standard DNN architectures without task-specific adaptations (like those in MTLDNNs) does not necessarily lead to better performance than traditional models, highlighting the importance of tailored deep learning approaches.

Key Results

  • Superior Prediction Accuracy: MTLDNN models (both the top single model and the top 10 ensemble) consistently outperformed six benchmark models, including NL-C, NL-NC, MNL-SPT, and DNN-SPT/JOINT.
  • ~5% Accuracy Gain over NL: The MTLDNNs achieved approximately 5% higher prediction accuracy on joint RP+SP, individual RP, and individual SP datasets compared to the NL models.
  • MTLDNNs outperform DNNs: The MTLDNNs also showed better performance than separate (DNN-SPT) and joint (DNN-JOINT) deep neural networks, underscoring the benefit of the multitask learning approach and its specific architectural and regularization designs.
  • Interpretability: The study successfully used gradient-based methods to interpret the MTLDNN models, providing insights into choice probabilities and variable elasticities, which are comparable to traditional econometric models.

Limitations and Future Work

  • High Model Complexity and Estimation Error: The paper acknowledges that the high complexity of DNNs, including MTLDNNs, can lead to larger estimation errors (variance), which might counteract the benefits of reduced approximation error (bias). This is a potential weakness that necessitates empirical validation.
  • Hyperparameter Tuning: The performance of MTLDNNs is highly dependent on hyperparameters, requiring extensive search (e.g., random search) to find optimal configurations.
  • Future Research Directions:
    • Explore more sophisticated MTLDNN architectures specifically designed for RP and SP data.
  • Investigate automatic ways to design new MTLDNN architectures using sequential modeling techniques.
  • Delve deeper into the relationship between MTLDNNs and NL models, given their visual architectural similarities.
  • Apply the MTLDNN framework to other domains and more complex scenarios (e.g., more than two tasks, panel data, heterogeneity, endogeneity).
  • Further research into how DNNs can address statistical concerns like heterogeneity and endogeneity.

Summary updated

BibTeX
@article{wang2020multitask,
  author = {Shenhao Wang and Qingyi Wang and Jinhua Zhao},
  title = {{Multitask learning deep neural networks to combine revealed and stated preference data}},
  journal = {Journal of Choice Modelling},
  year = {2020},
  doi = {10.1016/j.jocm.2020.100236},
}
Figure