Probabilistic Modelling for Unsupervised Analysis of Human Behaviour in Smart Cities

Yazan K. Qarout, Yordan P. Raykov, Max A. LittleSensors 2020 · DOI

TL;DR

This paper introduces the Adaptive Input Hidden Markov Model (AI-HMM), a novel probabilistic framework for analyzing complex human behavior in smart cities using sensor data. The AI-HMM excels at identifying dynamic behavioral trends and anomalies by adaptively incorporating contextual information, outperforming existing methods on real-world datasets.

What Problem Does This Paper Address?

Smart cities generate vast amounts of sensor data, offering a unique opportunity to understand human behavior for better urban planning, security, and resource management. However, analyzing this data, especially from sources like GPS trajectories, presents significant challenges. Existing methods often struggle with noisy, irregularly sampled data, and fail to account for crucial contextual factors like time of day or day of the week, which heavily influence behavior. Furthermore, many approaches rely on expensive, privacy-invading sensors like cameras, driving a need for more flexible and privacy-preserving alternatives. This paper aims to bridge this gap by developing a robust, unsupervised probabilistic model that can capture complex, context-dependent human behavior from low-dimensional sensor data.

Methods

The core contribution is the Adaptive Input Hidden Markov Model (AI-HMM). This model extends the traditional Hidden Markov Model (HMM) by incorporating an "adaptive input" layer. This layer allows the model to dynamically adjust its transition dynamics based on contextual information (represented by a semi-Markov variable, denoted as $\tau$). This $\tau$ can represent factors like "peak vs. off-peak" hours or "weekday vs. weekend" days.

Key aspects of the AI-HMM include:

  • Bayesian Nonparametric Approach: It utilizes Dirichlet Processes (DPs) to automatically infer the number of hidden states, avoiding the need to pre-specify model complexity.
  • Autoregressive Structure: It builds upon the Vector Autoregressive Hidden Markov Model (VAR-HMM) to capture the temporal dependencies within the observation sequences.
  • Adaptive Contextual Switching: The novel AI-HMM introduces an additional DP layer that allows for different base measures for different contextual realizations of $\tau$. This means the model can learn distinct transition dynamics for different contexts (e.g., weekday vs. weekend traffic) while still sharing a common underlying structure.
  • Inference: The model is trained using a blocked Gibbs sampling approach, which iteratively samples the model parameters, including the contextual input $\tau$ itself, allowing for adaptive inference.

Datasets

The proposed AI-HMM was evaluated on two distinct real-world datasets:

  1. Dodgers Loop Sensor Dataset: This dataset consists of loop sensor data counting vehicle passage at a freeway on-ramp in Los Angeles over 25 weeks. It exhibits clear weekday/weekend traffic patterns and the impact of baseball games. This dataset was used to model vehicular density counts, with features including vehicle count and hour of day.
  2. T-drive Dataset: This dataset comprises GPS trajectory data from over 10,000 taxis in Beijing collected over a week. It's characterized by noisy, sparse, and irregularly sampled data, making it a challenging benchmark for human mobility analysis. This dataset was used to model taxi trajectories, with features including time, longitude, latitude, hour of day, and average velocity.

Key Insights and Ideas

  • Context is King for Behavior: The research strongly emphasizes that human behavior is not static but highly dependent on context. Simply analyzing raw sensor data without considering factors like time of day or day type leads to a superficial understanding.
  • Adaptive Inference is Powerful: The ability of the AI-HMM to learn the contextual variable $\tau$ (e.g., taxi identity) from the data, rather than having it pre-assigned, significantly enhances its performance and provides deeper insights.
  • Unsupervised Learning for Complex Data: The AI-HMM demonstrates that sophisticated probabilistic models can effectively uncover meaningful patterns and behaviors in complex, noisy, and high-dimensional sensor data without requiring labeled examples.
  • Beyond Simple Clustering: Unlike methods like PCA, which can identify general trends, the AI-HMM can pinpoint specific behavioral states and their temporal dynamics, even identifying anomalies like noisy GPS points.

Key Results

  • Dodgers Loop Dataset: The AI-HMM successfully identified distinct states corresponding to morning/afternoon traffic peaks, evening baseball game traffic, and weekend traffic patterns. It outperformed PCA in identifying specific trends and associating them with particular days and times.
  • T-drive Dataset:
    • The AI-HMM identified six meaningful states for taxi trajectories, capturing patterns related to city center movement, highway travel, and airport waiting times.
  • When used for adaptive inference of taxi identity (using $\tau$ to represent taxi ID), the AI-HMM achieved 93% accuracy on unseen test data, demonstrating its ability to learn individual behavioral signatures.
  • Compared to a GRU autoencoder with K-means clustering, the AI-HMM provided more informative and distinct states that combined location, velocity, and time information more effectively.

Limitations and Future Work

  • Low-Dimensional Time-Series Data: The current AI-HMM is primarily designed for low-dimensional time-series data.
  • Future Work: The authors suggest extending the model to handle high-dimensional observation models and to incorporate multiple dependent contextual variables. They also propose exploring scenarios where the observation model can be high-dimensional.

Summary updated

BibTeX
@article{qarout2020probabilistic,
  author = {Yazan K. Qarout and Yordan P. Raykov and Max A. Little},
  title = {{Probabilistic Modelling for Unsupervised Analysis of Human Behaviour in Smart Cities}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20030784},
}
Figure