What are the ways of treatng missing values in XGboost?
naggar1 opened this issue · comments
Generally does the model performance get better with that ?
xgboost naturally accepts sparse feature format, you can directly feed data in as sparse matrix, and only contains non-missing value.
i.e. features that are not presented in the sparse feature matrix are treated as 'missing'. XGBoost will handle it internally and you do not need to do anything on it.
Internally, XGBoost will automatically learn what is the best direction to go when a value is missing. Equivalently, this can be viewed as automatically "learn" what is the best imputation value for missing values based on reduction on training loss.
I haven't done formal comparison with other methods, but I think it should be comparable, and it also gives computation benefit when your feature matrix is sparse
well - if values are not provided, it takes it as missing. So are all 0 values also treated as missing?
Example: A column has 25 values, 15 are 1, 5 are missing/NA and 5 are 0.
Are the 5 + 5 = 10 treated as missing?
It will depends on how you present the data. If you put data in as LIBSVM format, and list zero features there, it will not be treated as missing
it maybe extremely difficult to list 0 features in case of sparse data. So should we avoid xgboost in cases where there is missing data and many 0 features?
Just gave a quick glance of the code ( it is beautiful ,by the way), it is very interesting the way you treat the missing values - it is depending how to make the tree better. Does this method/algorithm have name?
Normally, it is fine that you treat missing and zero all as zero:)
On Sat, Aug 30, 2014 at 5:11 AM, rkirana notifications@github.com wrote:
it maybe extremely difficult to list 0 features in case of sparse data. So
should we avoid xgboost in cases where there is missing data and many 0
features?—
Reply to this email directly or view it on GitHub
https://github.com/tqchen/xgboost/issues/21#issuecomment-53956745.
Sincerely,
Tianqi Chen
Computer Science & Engineering, University of Washington
I invent the protocol and tricks by my self, maybe you can just call it
xgboost. The general algorithm however, fits into framework of gradient
boosting.
On Sat, Aug 30, 2014 at 8:56 AM, maxliu notifications@github.com wrote:
Just gave a quick glance of the code ( it is beautiful ,by the way), it is
very interesting the way you treat the missing values - it is depending how
to make the tree better. Does this method/algorithm has name?—
Reply to this email directly or view it on GitHub
https://github.com/tqchen/xgboost/issues/21#issuecomment-53962310.
Sincerely,
Tianqi Chen
Computer Science & Engineering, University of Washington
I am not surprised by the seed of xgboost but the score is better than sklearn-GBR. The trick of missing value might be one of the reasons.
Have you published any paper for the boosting algorithm you used for xgboost? Unlike random forest, I could not find many code for boosting with parallel algorithm - may need to improve my google skill though.
I didn't yet publish any paper describing xgboost.
For parallel boosting tree code, the only one I am aware of so far is
http://machinelearning.wustl.edu/pmwiki.php/Main/Pgbrt . You can try it out
and compare with xgb if you are interested
On Sat, Aug 30, 2014 at 9:40 AM, maxliu notifications@github.com wrote:
I am not surprised by the seed of xgboost but the score is better than
sklearn-GBR. The trick of missing value might be one of the reasons.Have you published any paper for the boosting algorithm you used for
xgboost? Unlike random forest, I could not find many code for boosting with
parallel algorithm - may need to improve my google skill though.—
Reply to this email directly or view it on GitHub
https://github.com/tqchen/xgboost/issues/21#issuecomment-53963590.
Sincerely,
Tianqi Chen
Computer Science & Engineering, University of Washington
A follow up question-
While I understand how XGboost handles missing values within discrete variables, I'm not sure how does it handle continues (numeric) variables.
Can you please explain?
For continuous features, a missing(default) direction is learnt for missing value data to go into, so when the data of the speficific value is missing, then it goes to the default direction
Thanks Tianqi.
And what about missing continuous features in generalized linear models?
Hi Tianqi
I am looking for an algo which does no imputation of the missing values internally and yet works .How does the xgboost work internally to handle missing values(can you drop in some basic idea) ?
see https://arxiv.org/abs/1603.02754 sec 3.4
Xgboost also works in the presence of categorical features ?We don't need to prepocess them (binarisation,etc.).For e.g my dataset has a feature called city which has values-"Milan","Rome","venice".Can I present them to xgboost without any preproccessing at all?
Tianqi,
I have question about xgb.importance function. When I run this and look at the Real Cover it seems as though if there is any missing data in a feature, the Real Cover is NA. Is there anyway to deal with this issue to get some co-occurence count for each split?
Rex
Hi Tianqi,
my processing pipeline include normalize the features before learning. Also, i have a lot of indicators which are missing and not zero for negative indication.
as a result, my normalized indicators are 0 for indicated values and missing for non-indicated values.
will xgboost handle such behavior properly? (changing non-exists features to 0 will cause problems...)
Thanks,
Nadav
@tqchen
You wrote:
Internally, XGBoost will automatically learn what is the best direction to go when a value is missing. Equivalently, this can be viewed as automatically "learn" what is the best imputation value for missing values based on reduction on training loss.
What about a case when the train set has not missing values, but the test has?