Open abhijeetmote opened 4 years ago
The paper is here: https://arxiv.org/abs/1912.00290
To be short, XGBOD is an extended version of xgboost but using unsupervised models to enrich the feature space to achieve a better result. In the paper itself, xgboost's performance is also compared.
Thanks @yzhao062 for your input. If I see the object of XGBClassifier and XGBOD it shows the following parameters tuned. So bu tuning the hyperparameters it is considered as an unsupervised model. Correct me if I am wrong?
XGBClassifier | HyperParameters | XGBOD |
---|---|---|
None | base_score | 0.5 |
None | booster | gbtree |
None | colsample_bylevel | 1 |
None | colsample_bynode | None |
None | colsample_bytree | None |
None | gamma | 0 |
None | gpu_id | None |
'gain' | importance_type | None |
None | interaction_constraints | None |
None | learning_rate | 0.1 |
None | max_delta_step | 0 |
None | max_depth | 3 |
None | min_child_weight | 1 |
nan | missing | None |
None | monotone_constraints | None |
100 | n_estimators | 100 |
None | n_jobs | 1 |
None | num_parallel_tree | None |
'binary:logistic' | objective | 'binary:logistic' |
None | random_state | 0 |
None | reg_alpha | 0 |
None | reg_lambda | 1 |
None | scale_pos_weight | 1 |
None | subsample | 1 |
None | tree_method | None |
None | validate_parameters | None |
None) | verbosity | None |
Internally anything else is changed or if I just use XGBClassfier with tuned parameters it should work same as XGBOD
Thanks, Abhijeet
@yzhao062, Actually the problem is when I use XGBOD it's giving the accuracy very well. but If I try to train using spark multiprocessing it is getting failed. Giving memory heap error.
On the other hand XGBClassifier is working with spark multiprocessing but accuracy sometimes is not that good.
what do you suggest to overcome the issue? Thanks, Abhijeet
Hi @yzhao062 ,
I have seen the code of XGBOD, it is using XGBClassifier internally(wrapper written using it) and the same default parameters. So what would be the difference if I use XGBOD over XGBClassifier?
I have seen when I use XGBOD the model size is more compare to XGBClassifier, what is happening internally in XGBOD
your inputs are appreciated. Thanks, Abhijeet