Various studies on the prevention of fire accidents have recently been conducted, but they remain in the form of summary analysis or statistical analysis by region. If we could know whether there was a fire accident at the level of each building, we would be able to prevent fire accidents more precisely by targeting buildings in the fire inspection area. This paper proposes a method of manifold learning for predicting a particular group of fire-hazard buildings. However, when unmodified raw data is used for clustering, it is not possible to know which feature is related to fire accidents, and moreover we have the curse-of-dimensionality problem. To solve these problems, we use the deep neural network (DNN) for supervised manifold learning to transform this raw data into low-dimensional space. The DNN has a simple structure consisting of only a fully-connected layer. Our proposed method shows better performance than conventional clustering methods that use the supervised dimensionality reduction. In addition, we investigated the characteristics of sub-clusters within a target cluster through hierarchical cluster analysis.
[1]
Han-joon Kim,et al.
Estimation of Individual Micro Data from Aggregated Open Data
,
2017,
ArXiv.
[2]
Geoffrey E. Hinton,et al.
Visualizing Data using t-SNE
,
2008
.
[3]
Jimmy Ba,et al.
Adam: A Method for Stochastic Optimization
,
2014,
ICLR.
[4]
Kilian Q. Weinberger,et al.
Densely Connected Convolutional Networks
,
2016,
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[5]
Duen Horng Chau,et al.
Firebird: Predicting Fire Risk and Prioritizing Fire Inspections in Atlanta
,
2016,
KDD.
[6]
Chong Wang,et al.
Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin
,
2015,
ICML.