Ariadna Albors Zumel
@ariadnaalbors.bsky.social
PhD student at the University of Trento & former visiting scholar at CMU | Interested in machine learning, deep learning, criminology, human mobility and crime forecasting 🕵🏻♀️
3/3 Further research could explore alternative area boundaries (e.g., hexagonal tessellations) and incorporate more nuanced mobility features, such as mobility flows, to improve precision.
📄 Read the full paper here (open access): link.springer.com/article/10.1...
📄 Read the full paper here (open access): link.springer.com/article/10.1...
September 23, 2025 at 12:23 PM
3/3 Further research could explore alternative area boundaries (e.g., hexagonal tessellations) and incorporate more nuanced mobility features, such as mobility flows, to improve precision.
📄 Read the full paper here (open access): link.springer.com/article/10.1...
📄 Read the full paper here (open access): link.springer.com/article/10.1...
2/3 🔑 Key findings:
• ConvLSTM outperforms traditional baselines
• Achieves high recall (few false negatives), but low precision (many false positives)
• Violent crimes benefit from longer look-back periods, property crimes from shorter ones
• Adding mobility features improves predictive performance
• ConvLSTM outperforms traditional baselines
• Achieves high recall (few false negatives), but low precision (many false positives)
• Violent crimes benefit from longer look-back periods, property crimes from shorter ones
• Adding mobility features improves predictive performance
September 23, 2025 at 12:20 PM
2/3 🔑 Key findings:
• ConvLSTM outperforms traditional baselines
• Achieves high recall (few false negatives), but low precision (many false positives)
• Violent crimes benefit from longer look-back periods, property crimes from shorter ones
• Adding mobility features improves predictive performance
• ConvLSTM outperforms traditional baselines
• Achieves high recall (few false negatives), but low precision (many false positives)
• Violent crimes benefit from longer look-back periods, property crimes from shorter ones
• Adding mobility features improves predictive performance
1/3 Together with @mtizzoni.bsky.social & @gcampedelli.bsky.social , we built a deep learning framework using ConvLSTMs that combines historical crime data, fine-grained human mobility data, and sociodemographic data to forecast crime 12h ahead at very small spatial scales across four U.S. cities.
September 23, 2025 at 12:19 PM
1/3 Together with @mtizzoni.bsky.social & @gcampedelli.bsky.social , we built a deep learning framework using ConvLSTMs that combines historical crime data, fine-grained human mobility data, and sociodemographic data to forecast crime 12h ahead at very small spatial scales across four U.S. cities.