GPS Trajectory Completion Using End-to-End Bidirectional Convolutional Recurrent Encoder-Decoder Architecture with...

Area of Science:

• Ubiquitous Computing
• Data Science

Background:

• Large GPS datasets are crucial for urban computing, but data quality issues like sparse and incomplete trajectories hinder performance.
• Existing methods for handling incomplete GPS data often rely on complex heuristics and require domain expertise.

Purpose of the Study:

• To develop a deep learning model for generating missing points in GPS trajectories.
• To enhance the model's performance using an attention mechanism.

Main Methods:

• Proposed a deep learning-based bidirectional convolutional recurrent encoder-decoder architecture.
• Integrated an attention mechanism between the encoder and decoder.
• Evaluated the model on the Microsoft Geolife dataset with varying grid resolutions and missing data lengths.

Main Results:

• The proposed model significantly reduced the average displacement error compared to state-of-the-art methods.
• Performance improvements were observed across different grid resolutions and lengths of missing GPS segments.

Conclusions:

• Deep learning, specifically the proposed encoder-decoder architecture with attention, is effective for reconstructing incomplete GPS trajectories.
• This approach offers a more robust and less domain-dependent solution for improving GPS data quality in urban computing.