Imputation of missing time-activity data with long-term gaps: A multi-scale residual CNN-LSTM network model.

Area of Science:

• Human Mobility Research
• Data Science
• Machine Learning

Background:

• Time-activity (TA) data offers granular insights into human mobility patterns.
• Long-term missing data in TA datasets limits their utility for comprehensive studies.
• Accurate imputation of TA data is crucial for advancing mobility research.

Purpose of the Study:

• To develop and evaluate a novel two-step imputation method for addressing long-term gaps in time-activity data.
• To enhance the reliability and completeness of TA data for human mobility studies.
• To improve the reconstruction of activity sequences, durations, and spatial extents from incomplete datasets.

Main Methods:

• A two-step imputation approach combining word2vec and a CNN-LSTM model.
• Utilizing the continuous bag-of-words word2vec model for efficient TA sequence representation.
• Employing a multi-scale residual Convolutional Neural Network (CNN)-stacked Long Short-Term Memory (LSTM) model for temporal dependency capture and prediction.

Main Results:

• The proposed method achieved an 84% prediction accuracy in imputing missing TA data.
• Demonstrated success in reconstructing the sequence, duration, and spatial extent of human activities.
• Validated using mobile phone-based TA data from 180 individuals with 10-fold cross-validation.

Conclusions:

• The developed two-step imputation method effectively addresses long-term gaps in TA data.
• The method shows high accuracy and reliability for reconstructing human mobility patterns.
• This approach offers a promising solution for utilizing incomplete TA data in scientific research.