Multi-Aspect Embedding for Attribute-Aware Trajectories

Thapana Boonchoo, Xiang Ao, Qing He

TL;DR

This paper introduces MAEAT, a novel method for learning trajectory representations that considers multiple "aspects" (like spatial location and time) simultaneously. This allows for a more nuanced understanding of trajectory similarity, outperforming existing methods in both accuracy and significantly faster execution times.

What Problem Does This Paper Address?

Modern GPS-enabled devices generate increasingly complex trajectory data, not just sequences of coordinates but also associated attributes like timestamps, geo-tags, or even user-generated content. Traditional trajectory similarity methods primarily focus on spatial proximity, failing to capture the rich semantic information embedded in these additional attributes. This limitation hinders accurate trajectory analysis for applications like urban planning or mobility pattern discovery. The paper aims to bridge this gap by developing a representation learning approach that can effectively integrate and leverage these multiple aspects of trajectory data.

Methods

The core of the proposed method, MAEAT (Multi-Aspect Embedding for Attribute-Aware Trajectories), is a representation learning technique inspired by sentence embedding. It treats trajectories as "sentences" and their multi-aspect attributes as "words." The key innovations are:

  • Multi-Aspect Embedding: MAEAT learns a unified embedding vector for each trajectory by considering multiple aspects (e.g., spatial and temporal) simultaneously.
  • Aspect Tokenization: Raw trajectory data is converted into discrete "aspect tokens." For spatial aspects, a Hidden Markov Model (HMM) is used for map matching to assign road segments. For temporal aspects, a density-based clustering method (mean shift) is applied to group timestamps into temporal "hotspots."
  • Regularization: A crucial component is an L2 regularization term that controls the importance of each aspect, preventing any single aspect from dominating the learning process and mitigating the "aspect imbalance problem." This ensures a balanced contribution from all considered aspects.

Datasets

The paper evaluates MAEAT on five benchmark datasets (i5sim, i5sim3, cross, cross2, cross3) and one real-world dataset of taxi trajectories from Porto. The benchmark datasets are noted to have ground truth for clustering tasks, while the Porto dataset is used for qualitative analysis.

Key Insights and Ideas

  • Beyond Spatial Similarity: The paper strongly advocates for moving beyond purely spatial similarity measures for trajectories, highlighting how temporal and other attributes can reveal crucial semantic connections.
  • Aspect Imbalance is Real: The authors identify and address a critical challenge in multi-aspect learning: the potential for one aspect to dominate others due to differences in data distribution (e.g., many more road segments than temporal clusters). Their regularization technique is key to managing this.
  • Embedding as a Unified Representation: MAEAT directly learns trajectory embeddings, avoiding potential information loss that can occur when aggregating embeddings of individual components (like locations).
  • Efficiency Matters: The research demonstrates that while capturing richer information, MAEAT also achieves significant speedups compared to existing methods, making it practical for large-scale applications.

Key Results

  • Quantitative Performance: MAEAT consistently outperforms state-of-the-art and baseline methods in clustering accuracy (measured by Purity and ARI) across multiple benchmark datasets. Even a simplified version considering only spatial aspects (MAEAT(S)) shows strong performance, with the full MAEAT model achieving superior results, especially on datasets with temporal variations.
  • Efficiency: MAEAT is remarkably efficient. It runs "two orders of magnitude" faster than compared methods in similarity computation. The training time also scales well with embedding size, and MAEAT is significantly faster than methods like T2Vec and LSTM.

Limitations and Future Work

  • Limited Aspects Explored: The current MAEAT implementation focuses on spatial and temporal aspects. The authors suggest extending the method to incorporate other attributes like images, text, or temperature.
  • Downstream Task Expansion: While MAEAT shows promise for mining and management tasks, the authors propose exploring its application in more specific downstream tasks such as popular route queries, trajectory search, and real-time route recommendations.

Summary updated

Abstract

Motivated by the proliferation of trajectory data produced by advanced GPS-enabled devices, trajectory is gaining in complexity and beginning to embroil additional attributes beyond simply the coordinates. As a consequence, this creates the potential to define the similarity between two attribute-aware trajectories. However, most existing trajectory similarity approaches focus only on location based proximities and fail to capture the semantic similarities encompassed by these additional asymmetric attributes (aspects) of trajectories. In this paper, we propose multi-aspect embedding for attribute-aware trajectories (MAEAT), a representation learning approach for trajectories that simultaneously models the similarities according to their multiple aspects. MAEAT is built upon a sentence embedding algorithm and directly learns whole trajectory embedding via predicting the context aspect tokens when given a trajectory. Two kinds of token generation methods are proposed to extract multiple aspects from the raw trajectories, and a regularization is devised to control the importance among aspects. Extensive experiments on the benchmark and real-world datasets show the effectiveness and efficiency of the proposed MAEAT compared to the state-of-the-art and baseline methods. The results of MAEAT can well support representative downstream trajectory mining and management tasks, and the algorithm outperforms other compared methods in execution time by at least two orders of magnitude.
BibTeX
@inproceedings{boonchoo2024multi,
  author = {Thapana Boonchoo and Xiang Ao and Qing He},
  title = {{Multi-Aspect Embedding for Attribute-Aware Trajectories}},
  booktitle = {Symmetry},
  year = {2024},
  url = {https://openreview.net/forum?id=4COlcB9aOJ},
}
Figure