Compare GWNBR with NB, some settings have worse results

[thekingofkings]

Table 4 abnormal observations

In Table 4 of the TKDE draft, comparing GWNBR with NB, some results are getting worse. Specifically, comparing column 2 vs 1. Adding POI feature makes NB better, meanwhile makes GWNBR worse.

[thekingofkings]

Reason

GWNBR relies on the bandwidth parameter h. Originally, we tune h under the D+G+P+T setting, and applies this tunned h for all the other feature settings in the same year. This is not right, because it is very likely that different feature settings requires different h to achieve optimal.