編集

次の方法で共有


rx_predict

Usage

revoscalepy.rx_predict(model_object=None, data=None, output_data=None, **kwargs)

Description

Generic function to compute predicted values and residuals using rx_lin_mod, rx_logit, rx_dtree, rx_dforest and rx_btrees objects.

Arguments

model_object

object returned from a call to rx_lin_mod, rx_logit, rx_dtree, rx_dforest and rx_btrees. Objects with multiple dependent variables are not supported.

data

a data frame or an RxXdfData data source object to be used for predictions. If a Spark compute context is being used, this argument may also be an RxHiveData, RxOrcData, RxParquetData or RxSparkDataFrame object or a Spark data frame object from pyspark.sql.DataFrame.

output_data

an RxXdfData data source object or existing data frame to store predictions.

kwargs

additional parameters

Returns

a data frame or a data source object of prediction results.

See also

rx_predict_default, rx_predict_rx_dtree, rx_predict_rx_dforest.

Example

import os
from revoscalepy import RxOptions, RxXdfData, rx_lin_mod, rx_predict, rx_data_step

sample_data_path = RxOptions.get_option("sampleDataDir")
mort_ds = RxXdfData(os.path.join(sample_data_path, "mortDefaultSmall.xdf"))
mort_df = rx_data_step(mort_ds)

lin_mod = rx_lin_mod("creditScore ~ yearsEmploy", mort_df)
pred = rx_predict(lin_mod, data = mort_df)
print(pred.head())

Output:

Rows Read: 100000, Total Rows Processed: 100000, Total Chunk Time: 0.058 seconds 
Rows Read: 100000, Total Rows Processed: 100000, Total Chunk Time: 0.006 seconds 
Computation time: 0.039 seconds.
Rows Read: 100000, Total Rows Processed: 100000, Total Chunk Time: Less than .001 seconds 
   creditScore_Pred
0        700.089114
1        699.834355
2        699.783403
3        699.681499
4        699.783403

Note: Function rx_predict does not run predictions but chooses the appropriate function rx_predict_default, rx_predict_rx_dtree, or rx_predict_rx_dforest based on the model which was given to it. Each of them has a different set of parameters described by their documentation.