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

Ting Zhong, Fang Liu, Fan Zhou, Goce Trajcevski, Kunpeng Zhang

Abstract

Social Circle Inference from Mobility data (SCIM) aims to reconstruct social networks by analyzing users' geographical footprints. However, existing methods often struggle with static location representations and independent friend predictions that ignore social correlations. This paper introduces SCIMAC (Social Circle Inference via self-Attention and Contextualized-embedding), a framework that leverages deep learning to capture the semantic nuances of human movement and the inherent dependencies within social circles. By employing contextualized Point of Interest (POI) embeddings and a label-correlated classification mechanism, SCIMAC significantly outperforms state-of-the-art baselines across multiple real-world datasets.

Introduction

In the era of location-based social networks (LBSNs), mobility data provides a rich source of information about human behavior and social structures. Inferring social ties from this data is a challenging task due to two primary hurdles: "context blindness" and "inference bias." Context blindness occurs when static embeddings fail to capture the varying meanings of a location based on the timing or sequence of visits. Inference bias arises when models predict potential friends independently, ignoring the statistical likelihood that social circles are highly correlated. SCIMAC addresses these challenges by treating mobility sequences similarly to natural language, allowing for dynamic, context-aware representations.

Methodology

The SCIMAC architecture is designed to handle the sequential and semantic complexity of human movement through three innovative layers:

  • Contextualized POI Embedding: Inspired by the ELMo architecture in Natural Language Processing, this layer uses a layered bi-directional LSTM to pre-train POI representations. Instead of a fixed vector for a specific location, the embedding changes based on the check-ins that precede and follow it in a trajectory, capturing the specific context of the visit.
  • Self-Attention Mechanism: Human movement is not a simple linear chain; some stops are more "socially informative" than others. The self-attention layer allows the model to weigh the importance of different sub-trajectories, capturing both short-term daily patterns and long-term personal preferences.
  • Label-Correlated Classification: To address inference bias, the authors implement a label embedding technique. During the decoding phase, the model considers a weighted average of the entire user label space rather than just picking the most likely friend. This allows the model to understand the hierarchy and overlap of social circles in a latent space.
The SCIMAC architecture, featuring the Encoder with contextualized embeddings and self-attention, and the Decoder with label-correlated classification.

The framework is evaluated on two primary tasks: SCIM-I, which identifies friends of a known user, and SCIM-II, which addresses the "cold-start" problem of identifying the social circle of an anonymous trajectory.

Datasets and Evaluation

The researchers evaluated SCIMAC on four real-world LBSN datasets containing both check-in histories and ground-truth social ties:

  • Brightkite and Gowalla: Large-scale datasets of location-based social interactions.
  • Tokyo and New York: Targeted datasets extracted from Foursquare, focusing on dense urban mobility patterns.

Key Results

SCIMAC consistently outperformed nearly ten state-of-the-art baselines, including TULER and DeepTSCI, across all datasets. On the Brightkite dataset for the SCIM-I task, SCIMAC achieved a Macro-F1 score of 0.8095, a significant improvement over the next best baseline (DeepTSCI at 0.7320).

Furthermore, the model demonstrated robust performance in the more difficult SCIM-II task, showing it can effectively generalize mobility patterns to users it has never seen during training. Ablation studies confirmed that removing any single component—contextualized embedding, self-attention, or label embedding—led to a drop in performance, validating the necessity of all three layers for robust social inference.

Conclusion and Future Work

SCIMAC represents a significant advancement in social circle inference by treating mobility as a polysemic sequence and accounting for global social context. By moving beyond static representations, the model achieves superior accuracy in reconstructing social networks from raw mobility data.

Despite its success, the authors identify several areas for future research. Incorporating additional data types, such as user profiles or text from posts, could further refine social context. Additionally, adding geographical constraints (e.g., friends typically living in the same city) could sharpen predictions for anonymous users. Finally, the authors emphasize the importance of developing privacy defense mechanisms to protect user anonymity against such high-accuracy inference techniques.

Summary updated

Abstract

Extracting knowledge from human mobility data is an important task for many downstream applications such as point-of-interest recommendation, motion trace identification, and personalized trip planning. A specific problem that has recently spurred research interest is the so-called Social Circle Inference from Mobility data (SCIM), aiming at inferring relationships among users based on mobility data and without any explicit structured network information. The existing methods either require partial social ties or fail to model the implicit correlations between user links, thereby suffering from critical inference bias. We present a novel SCIM framework, called SCIM via self-Attention and Contextualized-embedding (SCIMAC) - a methodology capturing multiple aspects of users' check-in behavior and complex motion patterns of different users. Instead of directly applying the recurrent model on training user trajectories, the proposed method introduces a new module for context-aware check-in representation learning by adaptively incorporating the internal states of the recurrent layers, which is more effective than the context-independent check-in embedding used in existing social circle inference approaches. To model the underlying correlations between labels, SCIMAC leverages a more sophisticated label embedding technique to adjust the penalties for correlated users, enabling a better understanding of the user's hierarchy in the label space, and alleviating the inference bias. We empirically demonstrate that our SCIMAC model significantly outperforms several state-of-the-art baselines on real-world datasets.
BibTeX
@article{zhong2019motion,
  author = {Ting Zhong and Fang Liu and Fan Zhou and Goce Trajcevski and Kunpeng Zhang},
  title = {{Motion Based Inference of Social Circles via Self-Attention and Contextualized Embedding}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2915535},
}
Figure