10/04/2026
Our recent work, โAddressing Long-Tailed Spatial and Category Imbalances in Citywide Incident Prediction", accepted in ๐๐๐๐ ๐๐ซ๐๐ง๐ฌ๐๐๐ญ๐ข๐จ๐ง๐ฌ ๐จ๐ง ๐๐ข๐ ๐๐๐ญ๐ (๐๐ฆ๐ฉ๐๐๐ญ ๐
๐๐๐ญ๐จ๐ซ: ๐.๐), focuses on improving citywide incident prediction under highly imbalanced real-world urban data. In smart city environments, incident records are often unevenly distributed across both locations and incident categories, where a few regions and frequent events dominate the dataset while many critical but rare incidents remain under-represented. This long-tailed nature makes conventional prediction models biased toward majority patterns and less effective for minority regions and rare incident types.
๐ ๐๐๐ซ๐ฅ๐ฒ ๐๐๐๐๐ฌ๐ฌ ๐ฅ๐ข๐ง๐ค: https://ieeexplore.ieee.org/document/11460224
๐๐จ๐ง๐ ๐ซ๐๐ญ๐ฎ๐ฅ๐๐ญ๐ข๐จ๐ง๐ฌ ๐ญ๐จ ๐๐ฅ๐ฅ ๐ญ๐ก๐ ๐๐ฎ๐ญ๐ก๐จ๐ซ๐ฌ:
Bhumika Chaudhary, PhD student at Department of Computer Science and Engineering, IIT Jodhpur
Dr. Debasis Das, Associate Professor at Department of Computer Science and Engineering, IIT Jodhpur
๐๐ฎ๐ฆ๐ฆ๐๐ซ๐ฒ:
Citywide incidents such as crimes, accidents, and public safety threats contribute to substantial societal disruption and economic loss. Accurate prediction of such incidents can significantly aid city administrators in proactive response planning. Existing approaches model the incident prediction as a spatio-temporal task but often neglect the inter-region spatial long-tailed distribution of incidents. This uneven distribution introduces spatial bias in learning, which causes models to overfit regions with frequent incidents (head regions) while underfitting the regions with occasional incidents (tail regions). Furthermore, model learning is hindered by intra-region category imbalance, where certain incident types (e.g., theft) dominate over rarer categories (e.g., robbery) within the same region. To address inter- and intra-region challenges, we propose an approach named SLIP (Spatial Long-tail Incident Prediction). Specifically, for inter-region skewness, SLIP adopts a multi-expert design comprising a common feature extraction backbone followed by three expert branches. In addition, to mitigate the intra-region category imbalance, we utilizes a variant of focal loss, particularly for positive-negative imbalance. SLIP outperforms spatio-temporal state-of-the-art methods by 2-11% in Macro F1, 4-11% in Micro F1, and 1-11% in Severity Weighted F1 across Los Angeles and Chicago cities for the urban crime dataset. Additionally, we incorporate fairness metrics into the evaluation and present a comprehensive comparison of spatio-temporal incident prediction.
![New acceptance at ๐๐๐๐ ๐๐ซ๐๐ง๐ฌ๐๐๐ญ๐ข๐จ๐ง๐ฌ ๐จ๐ง ๐๐จ๐ง๐ฌ๐ฎ๐ฆ๐๐ซ ๐๐ฅ๐๐๐ญ๐ซ๐จ๐ง๐ข๐๐ฌ. [IF: 10.9]๐๐ข๐ญ๐ฅ๐: "Secure IoT Communications with Optimized ...](https://img5.schoolandcollegelistings.com/623/881/865484766238812.jpg)
