pyspark randomForest feature importance: how to get column names from the column numbers

Question

I am using the standard (string indexer + one hot encoder + randomForest) pipeline in spark, as shown below

labelIndexer = StringIndexer(inputCol = class_label_name, outputCol="indexedLabel").fit(data)

string_feature_indexers = [
   StringIndexer(inputCol=x, outputCol="int_{0}".format(x)).fit(data)
   for x in char_col_toUse_names
]

onehot_encoder = [
   OneHotEncoder(inputCol="int_"+x, outputCol="onehot_{0}".format(x))
   for x in char_col_toUse_names
]
all_columns = num_col_toUse_names + bool_col_toUse_names + ["onehot_"+x for x in char_col_toUse_names]
assembler = VectorAssembler(inputCols=[col for col in all_columns], outputCol="features")
rf = RandomForestClassifier(labelCol="indexedLabel", featuresCol="features", numTrees=100)
labelConverter = IndexToString(inputCol="prediction", outputCol="predictedLabel", labels=labelIndexer.labels)
pipeline = Pipeline(stages=[labelIndexer] + string_feature_indexers + onehot_encoder + [assembler, rf, labelConverter])

crossval = CrossValidator(estimator=pipeline,
                          estimatorParamMaps=paramGrid,
                          evaluator=evaluator,
                          numFolds=3)
cvModel = crossval.fit(trainingData)

now after the the fit I can get the random forest and the feature importance using cvModel.bestModel.stages[-2].featureImportances, but this does not give me feature/ column names, rather just the feature number.

What I get is below:

print(cvModel.bestModel.stages[-2].featureImportances)

(1446,[3,4,9,18,20,103,766,981,983,1098,1121,1134,1148,1227,1288,1345,1436,1444],[0.109898803421,0.0967396441648,4.24568235244e-05,0.0369705839109,0.0163489685127,3.2286694534e-06,0.0208192703688,0.0815822887175,0.0466903663708,0.0227619959989,0.0850922269211,0.000113388896956,0.0924779490403,0.163835022713,0.118987129392,0.107373548367,3.35577640585e-05,0.000229569946193])

How can I map it back to some column names or column name + value format?
Basically to get the feature importance of random forest along with the column names.

Lumpkin

Abishek how did you do this in the end? – Alcmene

Answer 1

The transformed dataset metdata has the required attributes.Here is an easy way to do -

1. create a pandas dataframe (generally feature list will not be huge, so no memory issues in storing a pandas DF)

   pandasDF = pd.DataFrame(dataset.schema["features"].metadata["ml_attr"] 
   ["attrs"]["binary"]+dataset.schema["features"].metadata["ml_attr"]["attrs"]["numeric"]).sort_values("idx")
   

2. Then create a broadcast dictionary to map. broadcast is necessary in a distributed environment.

   feature_dict = dict(zip(pandasDF["idx"],pandasDF["name"])) 
   
   feature_dict_broad = sc.broadcast(feature_dict)
   

You can also look here and here

Magda

This should be the correct answer - it's concise and effective. Thank you! – Bergson

Answer 2

Hey why don't you just map it back to the original columns through list expansion. Here is an example:

# in your case: trainingData.columns 
data_frame_columns = ["A", "B", "C", "D", "E", "F"]
# in your case: print(cvModel.bestModel.stages[-2].featureImportances)
feature_importance = (1, [1, 3, 5], [0.5, 0.5, 0.5])

rf_output = [(data_frame_columns[i], feature_importance[2][j]) for i, j in zip(feature_importance[1], range(len(feature_importance[2])))]
dict(rf_output)

{'B': 0.5, 'D': 0.5, 'F': 0.5}

Boatbill

Yes, but you are missing the point that the column names changes after the stringindexer/ onehotencoder. The one which are combined by Assembler, I want to map to them. I sure can do it the long way, but I am more concerned whether spark(ml) has some shorter way, like scikit learn for the same :) – Lumpkin

Ah okay my bad. But yeh the long way should still be valid. I don't think there is short solution at the moment. The Spark ML API is not as powerful and verbose as the scikit learn ones. – Boatbill

Yeah I know :), just wanted to keep the question open for suggestions :). Thanks Dat – Lumpkin

Answer 3

I was not able to find any way to get the true initial list of the columns back after the ml algorithm, I am using this as the current workaround.

print(len(cols_now))

FEATURE_COLS=[]

for x in cols_now:

    if(x[-6:]!="catVar"):

        FEATURE_COLS+=[x]

    else:

        temp=trainingData.select([x[:-7],x[:-6]+"tmp"]).distinct().sort(x[:-6]+"tmp")

        temp_list=temp.select(x[:-7]).collect()

        FEATURE_COLS+=[list(x)[0] for x in temp_list]



print(len(FEATURE_COLS))

print(FEATURE_COLS)

I have kept a consistent suffix naming across all the indexer (_tmp) & encoder (_catVar) like:

column_vec_in = str_col

column_vec_out = [col+"_catVar" for col in str_col]



indexers = [StringIndexer(inputCol=x, outputCol=x+'_tmp')

            for x in column_vec_in ]


encoders = [OneHotEncoder(dropLast=False, inputCol=x+"_tmp", outputCol=y)

for x,y in zip(column_vec_in, column_vec_out)]



tmp = [[i,j] for i,j in zip(indexers, encoders)]

tmp = [i for sublist in tmp for i in sublist]

This can be further improved and generalized, but currently this tedious work around works best

Quod