Research map / Time series / Anomaly detection

TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond

NeurIPS 2026 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TraXion, a pre-training framework for multi-entity spatiotemporal event streams (MESES) combining a noise-detection objective, an entity-prototype contrastive objective, and a factorized transformer with a co-occurrence axis, evaluated on six mobility datasets, LANL auth logs, and eICU mortality.

Problem

Existing pre-training recipes for mobility and event streams import language-modeling objectives (next-token, masked-token) that ignore tuple-valued events, persistent entity signatures, and cross-entity co-occurrence structure.

Approach

Three axioms (joint-attribute density, entity-coupled representation, shared context) motivate a pre-training recipe pairing context/time perturbation detection with an InfoNCE loss on learned entity prototypes that also serve as input tokens, on a backbone with an added co-occurrence attention axis.

Key results

• Visit-level anomaly AP on NUMOSIM-LA reaches .2512 (TraXion) and .6020 agent-level (TraXion w/o co-occurrence) vs. best baseline ICAD .0283 visit-level and .0329 agent-level.
• Social link AP on Foursquare-Tokyo .9354 vs. strongest baseline H3GNN .9138, tops 8 of 9 columns.
• On LANL, TraXion w/o co-occurrence reaches event-level AP .0315 vs. LogBERT .0076 and AUROC .9402 vs. .8861.

Surprising findings

• Generic causal and masked-token pre-training (C4, C5) fall below no pre-training on LANL log AD and next-visit T±60.
• Removing the co-occurrence sub-layer at fine-tune improves several tasks (UA-Berlin/Atlanta, NUMOSIM agent-level, next-visit), so the co-occurrence axis is not uniformly beneficial.

Limitations

• Discriminative only; trails on next-visit time prediction (T±60) against TPP-style and Mobility-LLM baselines.
• Co-occurrence sub-layer is chosen per dataset on validation rather than learned.
• One checkpoint per dataset rather than a unified multi-instance MESES backbone.

Experimental setup

Datasets: NUMOSIM-LA Urban Anomalies Berlin Urban Anomalies Atlanta Foursquare-Tokyo Gowalla-Stockholm Gowalla-Austin LANL unified host-and-network authentication log PhysioNet eICU-CRD demo

Benchmarks: NUMOSIM-LA Urban Anomalies Berlin Urban Anomalies Atlanta Foursquare-Tokyo Gowalla-Stockholm Gowalla-Austin LANL eICU-CRD demo

Metrics: AP Max F1 AUROC Sens@Sp.9 H@10 H@20 MRR T±60

Baselines: CoBAD DeepBayesic USTAD ICAD UniTraj GETNext LoTNext MobTCast COAST TAPT TrajGPT Mobility-LLM

Builds on: ELECTRA InfoNCE Space2Vec Time2Vec UniTraj

Closest papers by meaning

• RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility (NeurIPS 2025 · similarity 0.85)
• Taming the Long Tail in Human Mobility Prediction (NeurIPS 2024 · similarity 0.84)
• Spatial-Temporal Interplay in Human Mobility: A Hierarchical Reinforcement Learning Approach with Hypergraph Representation (AAAI 2024 · similarity 0.83)
• Large Language Models as Urban Residents: An LLM Agent Framework for Personal Mobility Generation (NeurIPS 2024 · similarity 0.83)
• UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.82)
• Correlation-Attention Masked Temporal Transformer for User Identity Linkage Using Heterogeneous Mobility Data (AAAI 2025 · similarity 0.82)