ValueError: feature_names mismatch: in xgboost in the predict() function

Question

I have trained an XGBoostRegressor model. When I have to use this trained model for predicting for a new input, the predict() function throws a feature_names mismatch error, although the input feature vector has the same structure as the training data.

Also, in order to build the feature vector in the same structure as the training data, I am doing a lot inefficient processing such as adding new empty columns (if data does not exist) and then rearranging the data columns so that it matches with the training structure. Is there a better and cleaner way of formatting the input so that it matches the training structure?

Answer 1

This is the case where the order of column-names while model building is different from order of column-names while model scoring.

I have used the following steps to overcome this error

First load the pickle file

model = pickle.load(open("saved_model_file", "rb"))

extraxt all the columns with order in which they were used

cols_when_model_builds = model.get_booster().feature_names

reorder the pandas dataframe

pd_dataframe = pd_dataframe[cols_when_model_builds]

Answer 2

Try converting data into ndarray before passing it to fit/predict. For eg: if your train data is train_df and test data is test_df. Use below code:

train_x = train_df.values
test_x = test_df.values

Now fit the model:

xgb.fit(train_x,train_y)

Finally, predict:

pred = xgb.predict(test_x)

Hope this helps!

Answer 3

I also had this problem when i used pandas DataFrame (non-sparse representation).

I converted training and testing data into numpy ndarray.

          `X_train = X_train.as_matrix()
           X_test = X_test.as_matrix()` 

This how i got rid of that Error!

Answer 4

From what I could find, the predict function does not take the DataFrame (or a sparse matrix) as input. It is one of the bugs which can be found here https://github.com/dmlc/xgboost/issues/1238

In order to get around this issue, use as_matrix() function in case of a DataFrame or toarray() in case of a sparse matrix.

This is the only workaround till the bug is fixed or the feature is implemented in a different manner.

Answer 5

I came across the same problem and it's been solved by adding passing the train dataframe column name to the test dataframe via adding the following code:

test_df = test_df[train_df.columns]