Multi-task Generative Adversarial Network for Missing Mobility Data Imputation

What is it about?

Due to temporary failures of GPS equipment, the collected mobility data are typically incomplete. The incomplete mobility data will negatively affect existing LBSN-oriented studies, such as point-of-interest recommendation. Given this background, we propose a novel solution for missing mobility data imputation to tackle this challenge, aiming to identify missing records in user check-in sequences.

Featured Image

Photo by Tianyi Ma on Unsplash

Photo by Tianyi Ma on Unsplash

Why is it important?

Instead of straightforwardly modelling check-in sequences, we introduce coarse-grained spatial information, i.e., the corresponding region for each POI, to fully capture sequential dependencies. Then, we employ multi-task learning to train the generator by different tasks. The training of the auxiliary task helps the shared encoder to provide a better sequence representation, and assists the primary task to generate high-quality missing records.