Multiclass Classification with APL (Automated Predictive Library)

Common machine learning scenarios such as fraud detection, customer churn, employee flight risk, aim to predict Yes/No outcomes using binary classification models. But sometimes the target to predict has more than just two classes. This is the case of Delivery Timeliness that can have three categories: Early/On-time/Late.

From this article you will learn how to train and apply a multiclass classification model in a Python notebook with HANA ML APL.

The following example was built using HANA ML 2.12.220325 and APL 2209.

Census Income will be our training dataset.

from hana_ml import dataframe as hd

conn = hd.ConnectionContext(userkey='MLMDA_KEY')

sql_cmd =  """ 

select * from apl_samples.census 

where "marital-status" not in (

  select "marital-status" from apl_samples.census 

  group by "marital-status" having count(*) < 1500 )

order by "id"

"""

hfd_train = hd.DataFrame(conn, sql_cmd)

hfd_train.head(5).collect().style.hide_index()

Let’s check the size of the HANA dataframe in number of rows.

hfd_train.shape[0]

Marital status is our multiclass target.

col_key = 'id'

col_target = 'marital-status'

col_predictors = hfd_train.columns

col_predictors.remove(col_key)

col_predictors.remove(col_target)

col_predictors.remove('education-num')

len(col_predictors)

We do a fit and ask for a final model with no more than six variables. The processing is done within the HANA database.

from hana_ml.algorithms.apl.gradient_boosting_classification import GradientBoostingClassifier

apl_model = GradientBoostingClassifier()

apl_model.set_params(variable_selection_max_nb_of_final_variables = '6',

                     other_train_apl_aliases={'APL/VariableAutoSelection':'true'})

apl_model.fit(hfd_train, label=col_target, key=col_key, features=col_predictors)

The target distribution looks like this:

my_filter = "\"Partition\" = 'Estimation'"

df = apl_model.get_debrief_report('MultiClassTarget_Statistics').filter(my_filter).collect()

df.drop('Oid', axis=1, inplace=True)

df.drop('Target Key', axis=1, inplace=True)

format_dict = {'% Weight':'{:,.2f}%', 'Weight':'{:,.0f}'}

df.style.format(format_dict).hide_index()

At this point we choose to save the APL trained model.

from hana_ml.model_storage import ModelStorage

model_storage = ModelStorage(connection_context=conn, schema='USER_APL')

apl_model.name = 'My Multiclass Model'  

model_storage.save_model(model=apl_model, if_exists='replace')

model_storage.list_models()

One hour or one day later …

We are back. We load our multiclass model.

from hana_ml import dataframe as hd

conn = hd.ConnectionContext(userkey='MLMDA_KEY')

from hana_ml.model_storage import ModelStorage

model_storage = ModelStorage(connection_context=conn, schema='USER_APL')

apl_model = model_storage.load_model(name='My Multiclass Model')

apl_model.get_model_info()

We request the model reports and display first the accuracy overall.

from hana_ml.visualizers.unified_report import UnifiedReport

UnifiedReport(apl_model).build().display()

We take a deeper look with the class-by-class report:

Here are the variables that APL selected.

We want to know which variables were excluded during the training, and why:

df = apl_model.get_debrief_report('ClassificationRegression_VariablesExclusion').collect()

df = df[['Variable', 'Reason For Exclusion']]

df.style.hide_index()

We define a new dataframe with a few rows to try the APL model.

sql_cmd = 'select * from apl_samples.census where "id" between 550 and 554 order by "id"'

hfd_apply = hd.DataFrame(conn, sql_cmd)

hfd_apply.collect().style.hide_index()

We do a predict and ask for the top three reasons. Again, the processing is done within the database.

apl_model.set_params( extra_applyout_settings=

{ 'APL/ApplyExtraMode': 'Advanced Apply Settings', 

  'APL/ApplyPredictedValue': 'false', 

  'APL/ApplyProbability': 'false', 

  'APL/ApplyDecision': 'true', 

  'APL/ApplyReasonCode/TopCount': '3', 

  'APL/ApplyReasonCode/ShowStrengthValue': 'false', 

  'APL/ApplyReasonCode/ShowStrengthIndicator': 'false' }

)

df = apl_model.predict(hfd_apply).collect()

df.columns = ['Id', 'Actual', 'Prediction', 'Reason 1 Name', 'Reason 1 Value', 'Reason 2 Name', 'Reason 2 Value', 'Reason 3 Name', 'Reason 3 Value']

df.style.format({'Probability': '{:,.2%}'.format}).hide_index()

One can also request the score for each class. The class with the highest score becomes the prediction.

apl_model.set_params( extra_applyout_settings=

{ 'APL/ApplyExtraMode': 'Advanced Apply Settings', 

  'APL/ApplyPredictedValue': 'true', 

  'APL/ApplyProbability': 'false', 

  'APL/ApplyDecision': 'true', 

 }

)

df = apl_model.predict(hfd_apply).collect()

df.rename(columns={'TRUE_LABEL': 'Actual','PREDICTED': 'Prediction'}, inplace=True)

df.columns = [hdr.replace("gb_score_marital-status_", "") for hdr in df]

df.style.hide_index()

Another option is to export the model equation for scoring in stand-alone JavaScript.

apl_scoring_equation = apl_model.export_apply_code(code_type='JSON')

text_file = open("apl_model.json", "w")

text_file.write(apl_scoring_equation)

text_file.close()

Source: sap.com