BigML Bindings: 101 - Using a PCA

The PCA model is used to find the linear combination of your original features that best describes your data. In that sense, the goal of the model is to provide a transformation that allows dimensionality reduction. Following the schema described in the prediction workflow, document, this is the code snippet that shows the minimal workflow to create a PCA model and produce a single projection.

from bigml.api import BigML
# step 0: creating a connection to the service (default credentials)
# check how to set your credentials in the Authentication section
api = BigML()
# step 1: creating a source from the data in your local "data/iris.csv" file
source = api.create_source("data/iris.csv")
# waiting for the source to be finished. Results will be stored in `source`
api.ok(source)
# step 3: creating a dataset from the previously created `source`
dataset = api.create_dataset(source)
# waiting for the dataset to be finished
api.ok(dataset)
# step 5: creating a PCA model
pca = api.create_pca(dataset)
# waiting for the PCA to be finished
api.ok(pca)
# the input data to project
input_data = {"petal length": 4, "sepal length": 2, "petal width": 1,
              "sepal witdh": 3}
# getting the transformed components, the projection
projection = api.create_projection(pca, input_data)

If you want to configure some of the attributes of your PCA, like selecting a default numeric value, you can use the second argument in the create call.

# step 5: creating a PCA and using mean as numeric value when missing
pca = api.create_pca(dataset, {"default_numeric_value": "mean"})
# waiting for the PCA to be finished
api.ok(pca)

You can check all the available creation arguments in the API documentation.

If you want to add the generated principal components to the original dataset (or a different dataset), you can do so by creating a batch_projection resource. In the example, we’ll be assuming you already created a PCA following the steps 0 to 5 in the previous snippet and that you want to score the same data you used in the PCA model.

test_dataset = dataset
# step 10: creating a batch projection
batch_projection = api.create_batch_projection(pca, test_dataset)
# waiting for the batch_projection to be finished
api.ok(batch_projection)
# downloading the results to your computer
api.download_batch_projection(batch_projection,
                              filename='my_dir/my_projection.csv')

The batch projection output (as well as any of the resources created) can be configured using additional arguments in the corresponding create calls. For instance, to include all the information in the original dataset in the output you would change step 10 to:

batch_projection = api.create_batch_projection(pca, test_dataset,
                                               {"all_fields": True})

Check the API documentation to learn about the available configuration options for any BigML resource.

You can also project your data locally using the PCA class in the pca module. A simple example of that is:

from bigml.pca import PCA
local_pca = PCA("pca/6878ec46983efc21b000001b")
# Getting the projection of some input data
local_pca.projection({"petal length": 4, "sepal length": 2,
                      "petal width": 1, "sepal witdh": 3})

Or you could store first your PCA information in a file and use that file to create the local PCA object:

# downloading the anomaly detector JSON to a local file
from bigml.api import BigML
api = BigML()
api.export("pca/6878ec46983efc21b000001b",
           filename="my_pca.json")
# creating a PCA  object using the information in the file
from bigml.pca import PCA
local_pca = PCA("my_pca.json")
# getting the projection for some input data
local_pca.projection({"petal length": 4, "sepal length": 2,
                      "petal width": 1, "sepal witdh": 3})

If you want to get the projection locally for all the rows in a CSV file (first line should contain the field headers):

import csv
from bigml.pca import PCA
local_pca = PCA("pca/68714c667811dd5057000ab5")
with open("test_data.csv") as test_handler:
    reader = csv.DictReader(test_handler)
    for input_data in reader:
    # predicting for all rows
        print local_pca.projection(input_data)

Every modeling resource in BigML has its corresponding local class. Check the Local resources section of the documentation to learn more about them.