Conditional Generative Adversarial Networks (CGAN) for Mobility Data Fusion

This code implements the paper, Kim et al. (2021). Imputing Qualitative Attributes for Trip Chains Extracted from Smart Card Data Using a Conditional Generative Adversarial Network. Transportation Research Part C. Under Review.

Overview

This model aims to estimate the qualitative attributes of large-scale passively collected data (smart card data) using small-scale travel survey data, based on data fusion. The CGAN trains probability distribution of qualitative attributes given trip-chain attributes by mimicking the small-scale survey data..

Getting Started

Dependencies

• Python 3.6.10
• Tensorflow 2.4.1, Keras 2.4.3

Components

Dataset

• 'Data' only contains pertubated and sampled smart card and travel survey data due to limited permission.
• train/test_incomplete data indicate the smart card containing trip-chain attributes
• train/test_complete data indicate the travel survey containing trip-chain and qualitative attributes
• Other data is obtained from the DataPreprocessing.ipynb

DataPreprocessing.ipynb

• DataPreprocessing transforms the trip-chain attributes into sequential ndarray to use for Tensorflow
• Detailed descriptions are provided in the notebook files.

2D-Transformer.ipynb

• Step-by-step implementation of CGAN for mobility data fusion is provided
• Class for Transformer, 1D-Positional, and 2D-Locational encoding are defined
• The code include all parts in the paper: Model structure (2D-Transformer), Model training (Conditional WGAN-GP), Evaluation (Fidelity and Diversity), and Visualization
• Pretained model with full training data is provided in the 'Py_generator'

BERT_Embed.ipynb

• BERT transforms categorical qualitative attributes into numeric one to use for calculating precision and recall
• Pretained model is also provided ('MLM_Embed_indiv.h5')

Notice

• Full paper will be provided after the peer-review process
• Detail logic behind the code is described in the full paper

Authors

@Eui-Jin Kim

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

• WGAN-GP
• Transformer
• Evaluation
• 2D-locational Encoding