The art of using pipelines

Pipelines are a natural way to think about a machine learning system. Indeed with some practice a data scientist can visualise data "flowing" through a series of steps. The input is typically some raw data which has to be processed in some manner. The goal is to represent the data in such a way that is can be ingested by a machine learning algorithm. Along the way some steps will extract features, while others will normalize the data and remove undesirable elements. Pipelines are simple, and yet they are a powerful way of designing sophisticated machine learning systems.

Both scikit-learn and pandas make it possible to use pipelines. However it's quite rare to see pipelines being used in practice (at least on Kaggle). Sometimes you get to see people using scikit-learn's pipeline module, however the pipe method from pandas is sadly underappreciated. A big reason why pipelines are not given much love is that it's easier to think of batch learning in terms of a script or a notebook. Indeed many people doing data science seem to prefer a procedural style to a declarative style. Moreover in practice pipelines can be a bit rigid if one wishes to do non-orthodox operations.

Although pipelines may be a bit of an odd fit for batch learning, they make complete sense when they are used for online learning. Indeed the UNIX philosophy has advocated the use of pipelines for data processing for many decades. If you can visualise data as a stream of observations then using pipelines should make a lot of sense to you. We'll attempt to convince you by writing a machine learning algorithm in a procedural way and then converting it to a declarative pipeline in small steps. Hopefully by the end you'll be convinced, or not!

In this notebook we'll manipulate data from the Kaggle Recruit Restaurants Visitor Forecasting competition. The data is directly available through River's datasets module.

from pprint import pprint
from river import datasets

for x, y in datasets.Restaurants():
    pprint(x)
    pprint(y)
    break

Downloading https://maxhalford.github.io/files/datasets/kaggle_recruit_restaurants.zip (4.28 MiB)
Uncompressing into /home/runner/river_data/Restaurants
{'area_name': 'Tōkyō-to Nerima-ku Toyotamakita',
 'date': datetime.datetime(2016, 1, 1, 0, 0),
 'genre_name': 'Izakaya',
 'is_holiday': True,
 'latitude': 35.7356234,
 'longitude': 139.6516577,
 'store_id': 'air_04341b588bde96cd'}
10

We'll start by building and running a model using a procedural coding style. The performance of the model doesn't matter, we're simply interested in the design of the model.

from river import feature_extraction
from river import linear_model
from river import metrics
from river import preprocessing
from river import stats
from river import utils

means = (
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 7)),
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 14)),
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 21))
)

scaler = preprocessing.StandardScaler()
lin_reg = linear_model.LinearRegression()
metric = metrics.MAE()

for x, y in datasets.Restaurants():

    # Derive date features
    x['weekday'] = x['date'].weekday()
    x['is_weekend'] = x['date'].weekday() in (5, 6)

    # Process the rolling means of the target  
    for mean in means:
        x = {**x, **mean.transform_one(x)}
        mean.learn_one(x, y)

    # Remove the key/value pairs that aren't features
    for key in ['store_id', 'date', 'genre_name', 'area_name', 'latitude', 'longitude']:
        x.pop(key)

    # Rescale the data
    scaler.learn_one(x)
    x = scaler.transform_one(x)

    # Fit the linear regression
    y_pred = lin_reg.predict_one(x)
    lin_reg.learn_one(x, y)

    # Update the metric using the out-of-fold prediction
    metric.update(y, y_pred)

print(metric)

MAE: 8.316538

We're not using many features. We can print the last x to get an idea of the features (don't forget they've been scaled!)

pprint(x)

{'is_holiday': -0.23103573677646685,
 'is_weekend': 1.6249280076334165,
 'weekday': 1.0292832579142892,
 'y_mean_by_store_id': -1.3980979075298516}

The above chunk of code is quite explicit but it's a bit verbose. The whole point of libraries such as River is to make life easier for users. Moreover there's too much space for users to mess up the order in which things are done, which increases the chance of there being target leakage. We'll now rewrite our model in a declarative fashion using a pipeline à la sklearn.

from river import compose


def get_date_features(x):
    weekday =  x['date'].weekday()
    return {'weekday': weekday, 'is_weekend': weekday in (5, 6)}


model = compose.Pipeline(
    ('features', compose.TransformerUnion(
        ('date_features', compose.FuncTransformer(get_date_features)),
        ('last_7_mean', feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 7))),
        ('last_14_mean', feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 14))),
        ('last_21_mean', feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 21)))
    )),
    ('drop_non_features', compose.Discard('store_id', 'date', 'genre_name', 'area_name', 'latitude', 'longitude')),
    ('scale', preprocessing.StandardScaler()),
    ('lin_reg', linear_model.LinearRegression())
)

metric = metrics.MAE()

for x, y in datasets.Restaurants():

    # Make a prediction without using the target
    y_pred = model.predict_one(x)

    # Update the model using the target
    model.learn_one(x, y)

    # Update the metric using the out-of-fold prediction
    metric.update(y, y_pred)

print(metric)

MAE: 8.413859

We use a Pipeline to arrange each step in a sequential order. A TransformerUnion is used to merge multiple feature extractors into a single transformer. The for loop is now much shorter and is thus easier to grok: we get the out-of-fold prediction, we fit the model, and finally we update the metric. This way of evaluating a model is typical of online learning, and so we put it wrapped it inside a function called progressive_val_score part of the evaluate module. We can use it to replace the for loop.

from river import evaluate

model = compose.Pipeline(
    ('features', compose.TransformerUnion(
        ('date_features', compose.FuncTransformer(get_date_features)),
        ('last_7_mean', feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 7))),
        ('last_14_mean', feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 14))),
        ('last_21_mean', feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 21)))
    )),
    ('drop_non_features', compose.Discard('store_id', 'date', 'genre_name', 'area_name', 'latitude', 'longitude')),
    ('scale', preprocessing.StandardScaler()),
    ('lin_reg', linear_model.LinearRegression())
)

evaluate.progressive_val_score(dataset=datasets.Restaurants(), model=model, metric=metrics.MAE())

MAE: 8.413859

Notice that you couldn't have used the progressive_val_score method if you wrote the model in a procedural manner.

Our code is getting shorter, but it's still a bit difficult on the eyes. Indeed there is a lot of boilerplate code associated with pipelines that can get tedious to write. However River has some special tricks up it's sleeve to save you from a lot of pain.

The first trick is that the name of each step in the pipeline can be omitted. If no name is given for a step then River automatically infers one.

model = compose.Pipeline(
    compose.TransformerUnion(
        compose.FuncTransformer(get_date_features),
        feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 7)),
        feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 14)),
        feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 21))
    ),
    compose.Discard('store_id', 'date', 'genre_name', 'area_name', 'latitude', 'longitude'),
    preprocessing.StandardScaler(),
    linear_model.LinearRegression()
)

evaluate.progressive_val_score(datasets.Restaurants(), model, metrics.MAE())

MAE: 8.413859

Under the hood a Pipeline inherits from collections.OrderedDict. Indeed this makes sense because if you think about it a Pipeline is simply a sequence of steps where each step has a name. The reason we mention this is because it means you can manipulate a Pipeline the same way you would manipulate an ordinary dict. For instance we can print the name of each step by using the keys method.

for name in model.steps:
    print(name)

TransformerUnion
Discard
StandardScaler
LinearRegression

The first step is a FeatureUnion and it's string representation contains the string representation of each of it's elements. Not having to write names saves up some time and space and is certainly less tedious.

The next trick is that we can use mathematical operators to compose our pipeline. For example we can use the + operator to merge Transformers into a TransformerUnion.

model = compose.Pipeline(
    compose.FuncTransformer(get_date_features) + \
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 7)) + \
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 14)) + \
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 21)),

    compose.Discard('store_id', 'date', 'genre_name', 'area_name', 'latitude', 'longitude'),
    preprocessing.StandardScaler(),
    linear_model.LinearRegression()
)

evaluate.progressive_val_score(datasets.Restaurants(), model, metrics.MAE())

MAE: 8.413859

Likewhise we can use the | operator to assemble steps into a Pipeline.

model = (
    compose.FuncTransformer(get_date_features) +
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 7)) +
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 14)) +
    feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), 21))
)

to_discard = ['store_id', 'date', 'genre_name', 'area_name', 'latitude', 'longitude']

model = model | compose.Discard(*to_discard) | preprocessing.StandardScaler()

model |= linear_model.LinearRegression()

evaluate.progressive_val_score(datasets.Restaurants(), model, metrics.MAE())

MAE: 8.413859

Hopefully you'll agree that this is a powerful way to express machine learning pipelines. For some people this should be quite remeniscent of the UNIX pipe operator. One final trick we want to mention is that functions are automatically wrapped with a FuncTransformer, which can be quite handy.

model = get_date_features

for n in [7, 14, 21]:
    model += feature_extraction.TargetAgg(by='store_id', how=utils.Rolling(stats.Mean(), n))

model |= compose.Discard(*to_discard)
model |= preprocessing.StandardScaler()
model |= linear_model.LinearRegression()

evaluate.progressive_val_score(datasets.Restaurants(), model, metrics.MAE())

MAE: 8.413859

Naturally some may prefer the procedural style we first used because they find it easier to work with. It all depends on your style and you should use what you feel comfortable with. However we encourage you to use operators because we believe that this will increase the readability of your code, which is very important. To each their own!

Before finishing we can take an interactive look at our pipeline.

model

get_date_features

def get_date_features(x): weekday = x['date'].weekday() return {'weekday': weekday, 'is_weekend': weekday in (5, 6)}

y_mean_by_store_id

TargetAgg ( by=['store_id'] how=Rolling ( obj=Mean () window_size=7 ) target_name="y" )

y_mean_by_store_id

TargetAgg ( by=['store_id'] how=Rolling ( obj=Mean () window_size=14 ) target_name="y" )

y_mean_by_store_id

TargetAgg ( by=['store_id'] how=Rolling ( obj=Mean () window_size=21 ) target_name="y" )

~['area_name', [...]

Discard ( area_name date genre_name latitude longitude store_id )

StandardScaler

StandardScaler ( with_std=True )

LinearRegression

LinearRegression ( optimizer=SGD ( lr=Constant ( learning_rate=0.01 ) ) loss=Squared () l2=0. l1=0. intercept_init=0. intercept_lr=Constant ( learning_rate=0.01 ) clip_gradient=1e+12 initializer=Zeros () )