TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond

AI in healthcare
Published: arXiv: 2605.06906v1
Authors

Shang-Ling Hsu Mark Tenzer Cyrus Shahabi Khurram Shafique

Abstract

Human mobility differs from text and from generic time series in three structural ways: visits are tuple-valued events whose meaning depends on the joint distribution over location, time, and activity; users carry persistent signatures across trajectories; and visits are not independent across users, since co-location at shared places is a primary signal. Existing pre-training recipes for mobility import objectives from language modeling, treating trajectories as sentences and visits as tokens, an analogy that fails against each of the three properties above. These properties define a broader class, multi-entity spatiotemporal event streams (MESES), spanning enterprise authentication logs, electronic health records, and other event-stream domains where entities share infrastructure, schedules, or contexts. We make the properties precise as three axioms that any pre-training framework for MESES should satisfy, and introduce TraXion, whose objectives and architecture are jointly designed to meet them. A single TraXion checkpoint per dataset beats task-specific baselines on every task across six public mobility datasets covering anomaly detection, next-POI recommendation, next-visit prediction, and social-link prediction. The same recipe, applied unchanged to enterprise authentication logs and ICU mortality prediction, matches or exceeds prior work on both, showing that event streams from domains as different as mobility, security, and healthcare can be modeled under a single framework.

Paper Summary

Problem
Human mobility data, as well as other event-stream domains like security and healthcare, pose a unique challenge for machine learning models. Current pre-training frameworks, designed for language, fail to capture the structural properties of these data, such as visits being tuple-valued events, users carrying persistent signatures, and visits being coupled across users.
Key Innovation
The researchers introduce TraXion, a pre-training framework that satisfies three axioms: (1) events are tuple-valued, (2) entities carry persistent signatures, and (3) events are coupled across users. TraXion combines a noise-detection objective and an entity-prototype contrastive objective on a factorized Transformer backbone, which is designed to capture these properties.
Practical Impact
TraXion can be applied to various event-stream domains, such as mobility, security, and healthcare. The researchers demonstrate its effectiveness on six public mobility datasets, beating specialized task-specific baselines on every task. Additionally, TraXion outperforms prior work on enterprise authentication logs and ICU mortality prediction. This shows that TraXion can be a general-purpose framework for modeling event streams from different domains.
Analogy / Intuitive Explanation
Imagine you're at a busy coffee shop, and you see multiple people interacting with each other. Each person has their own unique signature, like a fingerprint, that distinguishes them from others. The events that occur, like ordering a coffee or chatting with a friend, are tuple-valued, meaning they have multiple components like location, time, and activity. TraXion is like a model that captures these structural properties of the data, allowing it to better understand and predict the behavior of individuals and groups in different event-stream domains.
Paper Information
Categories:
cs.LG
Published Date:

arXiv ID:

2605.06906v1

Quick Actions