You are on page 1of 10

DeepCrime

Prabhat Kanaujia : 2016CS50789


Sunil Kumar : 2016CS10314
Motivation
● Crime Prediction using Attentive Hierarchical RNN
● Starting point : DeepCrime: Attentive Hierarchical Recurrent Networks for
crime prediction, published in CIKM `18

● Explore the Temporal Dynamics, Category Dependencies, Interrelations with


Ubiquitous Data and also Unknown Temporal Dependencies.
DATA

200 OK
Sources
Real-world crime data collected from New York City (NYC) OpenData portal

24,031 POIs of 14 categories

311 Public Service Complaint Data

Source: https://data.cityofnewyork.us/
Data Preprocessing
Crime Categories: Robbery, Burglary, Felony Assault and Grand Larceny

Regions: across 77 disjoint regions in NYC

1. Crime Matrix, CM(i): for ith region, J(crime categories) * K(time intervals)
2. Urban Anomaly Matrix, AM(i): for ith region, J(crime categories) * K(time intervals)
3. Points-of-Interest Matrix: Spaces like Art/Shopping centres, etc. in a given
region: J(crime categories) * L(POI categories)
Source : Chao Huang, Junbo Zhang, Yu Zheng, and Nitesh V. Chawla. 2018. DeepCrime: Attentive Hierarchical Recurrent Networks for Crime
Prediction. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM '18). ACM, New York,
NY, USA, 1423-1432. DOI: https://doi.org/10.1145/3269206.3271793
Progress
1. Preprocessing datasets
2. Attention over LSTM is ready
3. Using Sklearn mlp classifier
Problems faced
1. Hyper parameters for full model are not mentioned
2. 311 service complaints are not actually used in specified model so need to
find a way to integrate
3. Dataset available is for 2006-2009 ( actual dataset is having corruption
issues)
Metrics
Macro F1 scores

Micro F1 scores

You might also like