[Paper Review] Traj2User: exploiting embeddings for computing similarity of users mobile behavior

Andrea Esuli, Lucas May Petry+2 more|arXiv (Cornell University)|Jul 31, 2018

Human Mobility and Location-Based Analysis8 references4 citations

TL;DR

Traj2User proposes a Word2Vec-inspired method to generate user embeddings from semantically enriched trajectories, learning similarity implicitly through neural network training. It outperforms baseline methods by 14.7% in mean reciprocal rank on user similarity tasks, demonstrating effective modeling of heterogeneous mobility semantics without explicit similarity function design.

ABSTRACT

Semantic trajectories are high level representations of user movements where several aspects related to the movement context are represented as heterogeneous textual labels. With the objective of finding a meaningful similarity measure for semantically enriched trajectories, we propose Traj2User, a Word2Vec-inspired method for the generation of a vector representation of user movements as user embeddings. Traj2User uses simple representations of trajectories and delegates the definition of the similarity model to the learning process of the network. Preliminary results show that Traj2User is able to generate effective user embeddings.

Motivation & Objective

• To address the challenge of computing user similarity from semantically rich, heterogeneous mobility data across space, time, and multiple semantic dimensions.
• To eliminate the need for manually defining similarity functions across diverse trajectory attributes such as transportation mode, weather, or activity.
• To learn a unified, compact vector representation (user embeddings) that captures complex user movement patterns from raw trajectory segments.
• To evaluate whether neural embedding learning can effectively model user similarity in real-world mobility data with limited semantic labels.

Proposed method

• Traj2User treats each trajectory segment as a 'sentence' and semantic labels as 'words', applying a skip-gram architecture inspired by Word2Vec to learn user embeddings.
• It uses negative sampling and stochastic gradient descent to train the embedding model, learning vector representations that encode semantic and behavioral similarities.
• The method processes segmented trajectories where each segment is labeled with multiple semantic attributes (e.g., mode of transport, weather, activity), forming a sequence of labels per user.
• User embeddings are generated by aggregating the learned vector representations of all segments associated with a given user.
• The model does not require explicit similarity functions or manual weighting of semantic dimensions; these are learned end-to-end from data.
• It leverages a custom synthetic dataset of 1,000 user pairs created by splitting real user movement descriptors to simulate similarity for evaluation.

Experimental results

Research questions

• RQ1 Can a neural embedding model effectively learn user similarity from semantically enriched trajectories without explicit design of similarity functions for each attribute?
• RQ2 How well does Traj2User perform compared to traditional methods like Sum, PPMI, SM, and their SVD variants in capturing user movement similarity?
• RQ3 Does the model generalize to detect groups of users with similar mobility patterns, even when labels are sparse or non-uniformly distributed?
• RQ4 Why do probabilistic methods like Softmax-based models perform poorly on this task despite success in NLP?
• RQ5 Can shorter embedding dimensions degrade performance gracefully, and does the model benefit from full-length embeddings on small datasets?

Key findings

• Traj2User achieves a 14.7% relative improvement in mean reciprocal rank (MRR) over the second-best method, Sum, on a 1,000-pair test set, with statistical significance (p = 3.66×10⁻⁵).
• The method outperforms traditional NLP-based techniques like PPMI, SM, and their SVD variants, which perform poorly despite success in language modeling tasks.
• Softmax-based methods are the worst performers, indicating that forcing a probabilistic interpretation of trajectory data is ill-suited for this task.
• Even with reduced embedding size (f=8), Traj2User maintains strong performance, outperforming Sum, indicating graceful degradation and potential for data compression.
• The model successfully captures within-group similarities in a simulated population of 2,000 users (20 groups of 100), visualizing consistent clustering of similar users via cosine similarity of embeddings.
• The results suggest that the distribution of label frequencies in the TagMyDay dataset (logarithmic, not Zipf-like) may contribute to the poor performance of standard NLP methods, though this requires further investigation.