search

thekingofkings / embedding

Dynamic graph embedding
MIT License
11 stars 7 forks source link

Use deep & wide combined regressor to predict crime #8

Closed thekingofkings closed 7 years ago

thekingofkings commented 7 years ago

The tensorflow-based linear regression model reproduces the same results as the baseline method in KDD16 paper.

For year 2012 at community area level, the MAE is 419.85 and MRE is 0.339.

To enhance the model, we use the embedding of non-linear features, together with the linear features, to learn a regressor.

Refer to the deep and wide learning tutorial.

thekingofkings commented 7 years ago

MAE is 803.8, and MRE is 0.6497.

The DNN embedding of non-linear features are not good. The reason is that the non-linear adjacency relationships are not sparse. And the sparse embedding method does not account for the propagation.