2
$\begingroup$

I am trying to use the XGBoost model to perform a multi-class classification over 40 classes.

The code is as follows:

xgb_params = list(colsample_bytree= 0.7,
                  subsample = 0.7,
                  eta = 0.05,
                  objective= 'multi:softmax',
                  max_depth= 5,
                  min_child_weight= 1,
                  eval_metric= "mlogloss", num_class = categoryclassnos,
                  nthread=4)

fit.xgb = xgb.train(params = xgb_params,
                    data = dtrain,
                    nrounds = 500,
                    watchlist = list(train = dtrain, test=dtest),
                    print_every_n = 50)

However, I am getting the following error:

Check failed: (info.labels.size()) != (0) label set cannot be empty

I have reproduced the dataset and the R script here.

Any help/ pointers are deeply appreciated.

$\endgroup$
  • $\begingroup$ In your code what does trainlabelsfactored <- as.integer(train$primarydeptt) - 1 return $\endgroup$ – grldsndrs Jun 1 '17 at 19:02
  • $\begingroup$ This returns an integer in the range [0,max] based on the factor values. XGBoost requires this range (starting from 0) for label. $\endgroup$ – Chints Jun 2 '17 at 1:45
  • $\begingroup$ Continuing to explore this... As an attempt, I added label value in dtest parameter. The error went away... I still don't get the logic as there is unlikely to be a label tag available for test data set while predicting.. $\endgroup$ – Chints Jun 2 '17 at 9:53
  • 1
    $\begingroup$ When creating the xgb.DMatrix (dtrain, dtest), specify the label parameter. $\endgroup$ – bradS Jun 27 '18 at 8:18
3
$\begingroup$

In python I needed to set the label param in dtrain = xgb.DMatrix(X_train,label=y_train,feature_names=cfg_col_X)

PS I was going to comment but missing rep points.

| improve this answer | |
$\endgroup$
1
$\begingroup$

To train a model you need two things, your training data(a matrix of variables and observations) and your labels (the known classification of each observation in the training data). Perhaps the last column in your data is actually the label? If so you will to vertically split your dtrain so that parameter data= all columns but the last, and a parameter label= just the last column.

Once you have the model, you can use it to predict on data that doesn’t have labels and the output will be the labels.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.