Preparación de datos

Machine learning con PySpark

Andrew Collier

Data Scientist, Fathom Data

¿Necesitas todas esas columnas?

+-----+-------+-------+------+----+----+------+------+----+-----------+
|maker|  model| origin|  type| cyl|size|weight|length| rpm|consumption|
+-----+-------+-------+------+----+----+------+------+----+-----------+
|Mazda|   RX-7|non-USA|Sporty|null| 1.3|  2895| 169.0|6500|       9.41|
|  Geo|  Metro|non-USA| Small|   3| 1.0|  1695| 151.0|5700|        4.7|
| Ford|Festiva|    USA| Small|   4| 1.3|  1845| 141.0|5000|       7.13|
+-----+-------+-------+------+----+----+------+------+----+-----------+

Elimina los campos maker y model.

Eliminar columnas

# O bien elimina las columnas que no quieres...
cars = cars.drop('maker', 'model')

# ... o selecciona las columnas que quieres conservar.
cars = cars.select('origin', 'type', 'cyl', 'size', 'weight', 'length', 'rpm', 'consumption')

+-------+------+----+----+------+------+----+-----------+
| origin|  type| cyl|size|weight|length| rpm|consumption|
+-------+------+----+----+------+------+----+-----------+
|non-USA|Sporty|null| 1.3|  2895| 169.0|6500|       9.41|
|non-USA| Small|   3| 1.0|  1695| 151.0|5700|        4.7|
|    USA| Small|   4| 1.3|  1845| 141.0|5000|       7.13|
+-------+------+----+----+------+------+----+-----------+

Filtrar datos faltantes

# ¿Cuántos valores faltan?
cars.filter('cyl IS NULL').count()

Elimina los registros con valores faltantes en la columna cylinders.

cars = cars.filter('cyl IS NOT NULL')

Elimina los registros con valores faltantes en cualquier columna.

cars = cars.dropna()

Modificar columnas

from pyspark.sql.functions import round

# Crea una nueva columna 'mass'
cars = cars.withColumn('mass', round(cars.weight / 2.205, 0))

# Convierte length a metros
cars = cars.withColumn('length', round(cars.length * 0.0254, 3))

+-------+-----+---+----+------+------+----+-----------+-----+
| origin| type|cyl|size|weight|length| rpm|consumption| mass|
+-------+-----+---+----+------+------+----+-----------+-----+
|non-USA|Small|  3| 1.0|  1695| 3.835|5700|        4.7|769.0|
|    USA|Small|  4| 1.3|  1845| 3.581|5000|       7.13|837.0|
|non-USA|Small|  3| 1.3|  1965| 4.089|6000|       5.47|891.0|
+-------+-----+---+----+------+------+----+-----------+-----+

Indexar datos categóricos

from pyspark.ml.feature import StringIndexer

indexer = StringIndexer(inputCol='type',
                        outputCol='type_idx')


# Asigna índices a las cadenas
indexer = indexer.fit(cars)


# Crea una columna con los índices
cars = indexer.transform(cars)

Usa stringOrderType para cambiar el orden.

+-------+--------+
|   type|type_idx|
+-------+--------+
|Midsize|     0.0| <- valor más frecuente
|  Small|     1.0|
|Compact|     2.0|
| Sporty|     3.0|
|  Large|     4.0|
|    Van|     5.0| <- valor menos frecuente
+-------+--------+

Indexar el país de origen

# Indexa el país de origen:
#
# USA     -> 0
# non-USA -> 1
#
cars = StringIndexer(
  inputCol="origin",
  outputCol="label"
).fit(cars).transform(cars)

+-------+-----+
| origin|label|
+-------+-----+
|    USA|  0.0|
|non-USA|  1.0|
+-------+-----+

Ensamblar columnas

Usa un vector assembler para transformar los datos.

from pyspark.ml.feature import VectorAssembler

assembler = VectorAssembler(inputCols=['cyl', 'size'], outputCol='features')

assembler.transform(cars)

+---+----+---------+
|cyl|size| features|
+---+----+---------+
|  3| 1.0|[3.0,1.0]|
|  4| 1.3|[4.0,1.3]|
|  3| 1.3|[3.0,1.3]|
+---+----+---------+

¡Vamos a practicar!

Machine learning con PySpark