Trajectory-User Linking via Variational AutoEncoder

Abstract

Trajectory-User Linking (TUL) is an essential task in Geo-tagged social media (GTSM) applications, enabling personalized Point of Interest (POI) recommendation and activity identification. Existing works on mining mobility patterns often model trajectories using Markov Chains (MC) or recurrent neural networks (RNN) -either assuming independence between non-adjacent locations or following a shallow generation process. However, most of them ignore the fact that human trajectories are often sparse, high-dimensional and may contain embedded hierarchical structures. We tackle the TUL problem with a semi-supervised learning framework, called TULVAE (TUL via Variational AutoEncoder), which learns the human mobility in a neural generative architecture with stochastic latent variables that span hidden states in RNN. TUL-VAE alleviates the data sparsity problem by leveraging large-scale unlabeled data and represents the hierarchical and structural semantics of trajectories with high-dimensional latent variables. Our experiments demonstrate that TULVAE improves efficiency and linking performance in real GTSM datasets, in comparison to existing methods.

Motion Based Inference of Social Circles via Self-Attention and Contextualized Embedding

“…(3) periodical pattern mining: finding (sub-)sequences and periodical motion patterns, enabling travel recommendation [2], life pattern understanding [35], recovering trajectories associated with users [3], [4] and next location prediction [46], [47].…”

“…Recently, deep learning techniques -especially ones based on recurrent neural networks (RNNs) such as Long-short Term Memory (LSTM) [26] and Gated Recurrent Units (GRU) [27] -have been widely used to capture the long term sequential influence and mobility patterns. Spatialtemporal RNN models [3], [46]- [48] extend the RNN model by incorporating temporal and spatial context in each time unit for various downstream tasks, such as trajectory classification [3], [4], POI recommendation [1], [49], [50] and prediction [46], [47]. However, these methods mostly focus on capturing the transition dependencies among POIs -and they neither explicitly model users' mobility similarity nor infer their social interactions.…”

“…The past decade has witnessed a rapid growth of both academic and practical interest on mining human mobility patterns from location based social networks (LBSN) such as Twitter, Foursquare and Weibo. Availability of large volumes of LBSN data has spurred research in studying user behavior and movement patterns that can be analyzed for various LBSN services such as point-of-interest (POI) recommendation [1], tour scheduling [2], associating users with specific trajectories [3], [4], etc.…”

DeepTrip

“…Deep generative models, such as Variational Auto-encoder (VAE) [10] and Generative Adversarial Networks (GANs) [8], have been widely used in computer vision and natural language processing. VAE can capture the latent variability from complex high dimensional data and have been successfully used to tackle trajectory classification problem [21] and friendship inference from human mobility [5,20]. GAN received broad attention due to the ability of generating high-quality image and fluent conversations.…”

A Fast Lightweight Spatiotemporal Activity Prediction Method

“…The recent studies [16], [36]- [38] use GTSM to predict user activity. Chong and Lim [39], Zhou et al [40] and Zhou et al [41] reveal the relationship between human behavior and temporal location by analyzing GTSM. S Vosoughi et al [42] learns tweet embeddings by CNN-LSTM endcoder-decoder from GTSM.…”