Cross-Interaction Hierarchical Attention Networks for Urban Anomaly Prediction

Chao Huang, Chuxu Zhang, Peng Dai, Liefeng Bo

Introduction

Predicting urban anomalies—such as traffic accidents, illegal parking, or noise complaints—is a cornerstone of effective smart city management. Anticipating these events allows for proactive resource deployment and improved public safety. However, urban anomaly prediction is a "wicked" problem characterized by two primary challenges: dynamic temporal dependencies and complex multi-dimensional interactions. Anomaly patterns are not static; they evolve seasonally and depend on the intricate "cross-talk" between geographical regions, specific time slots, and the category of the anomaly itself.

To address these challenges, the Cross-Interaction Hierarchical Attention Network (CHAT) framework was developed. CHAT leverages bidirectional recurrent networks and a two-tier hierarchical attention mechanism to capture long-term global temporal correlations and fuse spatial, temporal, and categorical signals effectively.

The CHAT Framework

The CHAT architecture is designed to process sequences of historical anomaly data through three integrated components:

  1. Temporal Dynamic Encoder: Utilizing a Bidirectional LSTM (BiLSTM), this component processes historical data by looking both forward and backward in time. This dual perspective captures a richer contextual representation of how urban anomalies evolve over time compared to standard unidirectional models.
  2. Temporal-wise Gating Module: Recognizing that not all historical events are equally predictive, this first layer of attention recalibrates the importance of different time steps. It enables the model to focus on the most relevant past moments for a given prediction.
  3. Interaction Attention Module: This core innovation creates embeddings for regions, time slots, and anomaly categories. It calculates pairwise interactions (Region-Time, Region-Category, and Time-Category) and uses an interaction-level attention mechanism to determine which cross-modal relationships are most critical for the current prediction.
The CHAT architecture, featuring the Temporal Dynamic Encoder and Hierarchical Attention Networks.

The model is optimized using a cross-entropy loss function to predict the likelihood of an anomaly occurring within a specific (region, category, time) triplet.

Data and Evaluation

The framework was evaluated using a massive real-world dataset from the New York City (NYC) 311 online reporting platform, spanning the entirety of 2014. The study focused on four key citywide anomalies: Blocked Driveway, Building Use, Noise, and Illegal Parking.

To test the model's robustness across different spatial resolutions, the authors partitioned NYC into two scales: a high-level scale consisting of 77 political districts and a fine-grained scale comprising 862 regions based on road segments.

Heatmap visualization of urban anomalies across New York City, illustrating the spatial complexity of the dataset.

Key Insights and Results

The experimental results demonstrate that CHAT consistently outperforms state-of-the-art baselines, including traditional time-series models (ARIMA), Bayesian methods (BIAP), and other deep learning architectures (ST-RNN, DRN).

  • Hierarchical Advantage: By treating temporal relevance and cross-dimensional interactions as distinct hierarchical layers, CHAT mimics human prioritization of information. This approach proves superior to flattening spatial and temporal data into a single vector.
  • Resilience to Sparsity: In fine-grained mapping scenarios where data is often sparse, CHAT’s joint embedding of all signals allows it to maintain high accuracy, whereas traditional models degrade significantly.
  • Performance Gains: CHAT achieved a notable 39.1% gain in F1-score over the next best deep learning models in the "Building Use" category, which is traditionally difficult to predict.
  • Ablation Success: Removing either the temporal gating or the interaction attention modules led to a significant drop in performance, confirming that both temporal context and cross-dimensional interactions are vital for accurate prediction.

Conclusion and Future Directions

The CHAT framework represents a significant advancement in urban anomaly prediction by effectively modeling the complex interactions between time, space, and event types. While powerful, future work is needed to adapt the model for real-time streaming data, allowing for on-the-fly parameter inference in city command centers. Additionally, applying the CHAT architecture to other urban sensing domains, such as air quality or energy consumption, could further validate its hierarchical interaction logic.

Summary updated

Abstract

Predicting anomalies (e.g., blocked driveway and vehicle collisions) in urban space plays an important role in assisting governments and communities for building smart city applications, ranging from intelligent transportation to public safety. However, predicting urban anomalies is not trivial due to the following two factors: i) The sequential transition regularities of anomaly occurrences is complex, which exhibit with high-order and dynamic correlations. ii) The Interactions between region, time and anomaly category is multi-dimensional in real-world urban anomaly forecasting scenario. How to fuse multiple relations from spatial, temporal and categorical dimensions in the predictive framework remains a significant challenge. To address these two challenges, we propose a Cross-Interaction Hierarchical Attention network model (CHAT) which uncovers the dynamic occurrence patterns of time-stamped urban anomaly data. Our CHAT framework could automatically capture the relevance of past anomaly occurrences across different time steps, and discriminates which types of cross-modal interactions are more important for making future predictions. Experiment results demonstrate the superiority of CHAT framework over state-of-the-art baselines.
BibTeX
@inproceedings{huang2020cross,
  author = {Chao Huang and Chuxu Zhang and Peng Dai and Liefeng Bo},
  title = {{Cross-Interaction Hierarchical Attention Networks for Urban Anomaly Prediction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/601},
}
Figure