ICAD: A Self-Supervised Autoregressive Approach for Multi-Context Anomaly Detection in Human Mobility Data�

​

Bita Azarijoo¹, Maria Despoina Siampou¹, John Krumm¹, Cyrus Shahabi¹

¹University of Southern California (USC)

​

SIGSPATIAL 2025

GPS Trajectory Vs Visit Sequences

GPS Trajectory

Sequence of visits

10:05 AM

10:15 AM

2

|

What is Anomaly Detection in Human Mobility?

​

​

​

​

Spatial Anomaly

Regular Routine

Temporal Anomaly

3

|

Why Visit-level Anomaly Detection is important?

7 PM-7 AM

8 AM-12 PM

8 AM-11:30 PM

4

|

Related Work

Trajectory anomaly detection mostly studied for taxi trajectories.

5

|

Related Work

6

|

Contributions

• ICAD (Interpretable Component-wise Anomaly Detection)

​

• It estimates anomaly scores for spatial and temporal components of a visit relative to normal patterns.

​

• For temporal signals, we propose a novel mode-margin scoring in continuous space.

​

• Interpretability through component-wise scoring.

7

|

Problem Formulation

Input:

• Sequence of visits
• location (latitude, longitude)
• Arrival time
• Departure Time

Approach:

• Next-visit prediction by TrajGPT¹

​

¹Hsu, Shang-Ling, et al. "Trajgpt: Controlled synthetic trajectory generation using a multitask transformer-based spatiotemporal model." Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 2024.

8

|

Background: TrajGPT

​

Visit Sequence� Encoding

Next Visit Prediction

Sequence Embedding

Transformer�Encoder

causal self-attention

Multihead CrossAttn

6 PM

Multihead CrossAttn

8 PM

9

|

Relative Anomaly Scoring

​

​

• Region Score (Discrete Space)

​

​

​

​

• Average deviation from the top-k predicted regions.

10

|

Relative Anomaly Scoring

Anomaly score < 1

• Anomalies are captured relative to high probable normal behaviors.

vs

Lower Likelihood Higher Distance

Higher Likelihood Lower Distance

• Arrival and Departure Score (Continuous Space)

​

Anomaly score ~ 50

11

|

Relative Anomaly Scoring

Component-wise Anomaly Detection

​

12

|

Experimental Setup

• Dataset: NUMOSIM [GeoAnomalies ‘24]
• Synthetic dataset.
• Los Angeles
• 2 months in 2024.
• # visit-records ～ 17.5 M

​

• Granularity Level:
• Visit-level
• Agent-level (aka sequence-level): maximum aggregation of visit-level anomaly scores.

​

• Evaluation Metrics:
• AP and AUROC.

13

|

Experimental Results

• Predicting all visit components (region, arrival time, departure time) capture fine-grained spatial and temporal anomalies.

​

• ICAD scores visits relative to high probable normal behavior in logarithmic space.

​

• Modeling temporal signals in continuous space helps fine-grained, but crucial deviation compared to others which usually discretize time.

14

|

Interpretation of Anomaly Scores

• Spatial Anomalies:
• Region score is the determining factor.
• Timing of spatial anomalies is atypical�compared to normal behavior.
• Temporal Anomalies:
• Negative region score SHAP values�push them towards normal behavior.
• Arrival and departure time contribute�the most to identifying a temporal�anomaly as anomalous.

​

15

|

Summary

• Granularity of Analysis: visit- & agent-level.
• Contribution: interpretability through relative anomaly scoring.
• Result: significant performance improvement in AP and AUROC.

Code

Paper

16

|

Thank You!� Any Questions? ☺

The End

17

|

Relative Multi-context Anomaly Scoring

Anomaly score ~ 50

Anomaly score ~ 1

vs

Lower Likelihood Higher Distance

Higher Likelihood Lower Distance

• Arrival and Departure Score (Continuous Space)

​

18

|

Experiment 2: Ablation on ST Components

1. Worst Performance of W/O Arrival:
1. Humans usually tend not to deviate significantly from their usual temporal patterns when visiting a regular location.
2. Combining Spatiotemporal signals is helpful.
• W/O weighted fusion in 2nd place in 3 out of 4 settings.

Adding anomaly scores.

19

|

Experiment 2: Temporal Signal Variations

20

|

Experiment 3: Relative Anomaly Scoring

• CDF-based:

​

​

​

• Relative Scoring vs CDF-based:

​

​

​

​

• Relative Scoring vs NLL:
• 3% improvement in AP and 2% is considerable in extremely rare anomaly detection problem.

​

21

|