project specification

people order food on a online e-commerce platform. drivers take food from restaurant(placeId) to the customer (delivery city and delivery detail)

target: predicts the time from placing an order to order completation.

approaches

feature analysis and engineer

• barplot, violinplot, boxplot for category vs numeric relation study
• heatmap for category,category vs numeric relation study
• kdeplot, histplot for distribution study
• statistics by group by operation to give more detail information
• openfe tool for more feature space searching

model study

5 model were studied.

model name	platform	score (rmse)
xgboost	sklearn	10.682, 10.60(by openfe augment)
lightgbm	sklearn	9.925
tabnet	torch	10.338
ft transformer	torch	13.36
resnet	tensorflow	10.23

end to end model

feature engineering by sklearn pipeline and end to end model was produced.

conclusion

• lightgbm is the best， tabnet and tesnet are also powerful
• statistics by group by operation make a good performance improvement.
• open fe tool make improvement to xgboost, but no improvement to other model.
• pretraining have better performance on tabnet
• ft transformer is not fit for such a problem