algorithm_lab1

A Rating Task

Reference

1. Use xgboost with grid search and cross validation. https://jessesw.com/XG-Boost/
2. XGBoost parameter settings. http://blog.csdn.net/wzmsltw/article/details/50994481

Some Results:

• Train data average: 4.01102555

• Compare different models:

model	parameters	rmse
knn	default	4.126
linear-svm	default	3.936
xgb-regressor	default	3.726

• Compare different targets:

target	model	rmse
orig	xgb	3.726
sqrt	xgb	3.973
log	xgb	4.445

• Test cascade prediction:

Totally, there are 27999 instances for training, while 2564 of them has target value >= 10.

So, it's very imbalanced. We use xgboost to predict whether the target value is >= 10. But only 17 of 11999 instances are predicted as true on the dev set. And 42 of 27999 instances are predicted as true on the train set. On dev set, the precision score of the prediction is 0.588. With cascade prediction, the rmse is 3.7255, slightly better than the rmse without it, 3.7258.