Closed thekingofkings closed 8 years ago
Truly observe that by learning two separate models for south and north Chicago will give better prediction accuracy.
In the commit above, the separate models do not have better performance.
One issue could be that when I separate model, I also separate the data. With less correlation information, the degenerated prediction accuracy is as expected.
Build separate models for north and south. Meanwhile, keep all the social/spatial lag information at hand. Namely, when train a model on North Chicago, the flow from south should be captured in the model.
Good examples are found. This thread should be closed. We move to issue #10
Motivation
In different regions, the feature effects are different. For example, in the South and North part of Chicago, the linear model have different parameters.
Solution
Build a dynamic coefficient model for the crime count prediction problem.
dynamic-coefficient-model.pdf
Target Here
In this thread, we focus the first issue. The motivating example to train dynamic coefficient model, instead of a single group of coefficients.
The dynamic coefficient model will be splitted into another thread.