Code Tutorial

Getting Started with GetML

JK
Dr. Johannes King on

Curious how automated feature engineering with getML works? This guide demonstrates getML's capabilities from a user perspective using a simple data set.

In this example, you will learn about the basic concepts of getML. You will tackle a simple problem using the Python API in order to gain a technical understanding of the benefits of getML. More specifically, you will learn how to do the following:

  1. Start a new project
  2. Define a data model
  3. Train a ML model

You have not installed getML on your machine yet? Head over to the installation instructions before you get started.

Introduction

Automated machine learning (AutoML) has attracted a great deal of attention in recent years. The goal is to simplify the application of traditional machine learning methods to real world business problems by automating key steps of a data science project, such as feature extraction, model selection, and hyperparameter optimization. With AutoML, data scientists are able to develop and compare dozens of models, gain insights, generate predictions, and solve more business problems in less time.

While it is often claimed that AutoML covers the complete workflow of a data science project - from the raw data set to the deployable machine learning models - current solutions have one major drawback: They cannot handle real world business data. This data typically comes in the form relational data. The relevant information is scattered over a multitude of tables that are related via so-called join keys. In order to start an AutoML pipeline, a flat feature table has to be created from the raw relational data by hand. This step is called feature engineering and is a tedious and error-prone process that accounts for up to 90% of the time in a data science project.

getML adds automated feature engineering on relational data and time series to AutoML. The getML algorithms, Multirel and Relboost, find the right aggregations and subconditions needed to construct meaningful features from the raw relational data. This is done by performing a sophisticated, gradient-boosting-based heuristic. In doing so, getML brings the vision of end-to-end automation of machine learning within reach for the first time. Note that getML also includes automated model deployment via a HTTP endpoint or database connectors. This topic is covered in other material.

All functionality of getML is implemented in the so-called getML engine. It is written in C++ to achieve the highest performance and efficiency possible and is responsible for all the heavy lifting. The getML Python API acts as a bridge to communicate with engine. In addition, the getML monitor provides a Go-based graphical user interface to ease working with getML and significantly accelerate your workflow.

In this article you will learn the basic steps and commands to tackle your data science projects using the Python API. For illustration purpose we will also touch how an example data set like the one used here would have been dealt with using classical data science tools. In contrast, we will show how the most tedious part of a data science project - merging and aggregating a relation data set - is automated using getML. At the end of this tutorial you are ready to tackle your own use cases with getML or dive deeper into our software using a variety of follow-up material.

Starting a new project

After you’ve successfully installed getML, you can launch it by executing the getml-cli command line interface or double-clicking the application icon. This launches both the getML engine and monitor.

Before diving into the actual project, you need to log into the getML suite. This happens in the getML Monitor, the frontend to the engine. If you open the browser of your choice and visit http://localhost:1709/, you’ll see a login screen. Click ‘create new account’ and follow the indicated steps. After you’ve activated your account by clicking the link in the activation E-mail you’re ready to go. From now on, the entire analysis is run from Python. We will cover the getML monitor in a later tutorial but feel free to check what is going on while following this guide.

import getml
print("getML version: {}".format(getml.__version__))
getML version: 0.10.0

First, we create a new project. All data sets and models belonging to a project will be stored in ~/.getML/getml-VERSION/projects.

getml.engine.set_project('getting_started')
Creating new project 'getting_started'

Data set

The data set used in this tutorial consists of 2 tables. The so-called population table represents the entities we want to make a prediction about in the analysis. The peripheral table contains additional information and is related to the population table via a join key. Such a data set could appear for example in a customer churn analysis where each row in the population table represents a customer and each row in the peripheral table represents a transaction. It could also be part of a predictive maintenance campaign where each row in the population table corresponds to a particular machine in a production line and each row in the peripheral table to a measurement from a certain sensor.

In this guide, however, we do not assume any particular use case. After all, getML is applicable to a wide range of problems from different domains. Use cases from specific fields are covered in other articles.

population_table, peripheral_table = getml.datasets.make_numerical(
    n_rows_population=500,
    n_rows_peripheral=100000,
    random_state=1709
)

# Save data frames to disk in order to make them available after restarting the engine
population_table.save()
peripheral_table.save()

This is the resulting population table

  join_key column_01 targets time_stamp
0 0 -0.629518 101.0 1970-01-01 11:18:00.114278400
1 1 -0.962169 88.0 1970-01-01 21:35:41.185276800
2 2 0.732649 17.0 1970-01-01 02:03:27.430905600
3 3 -0.462678 74.0 1970-01-01 08:45:55.322755200
4 4 -0.837399 96.0 1970-01-01 10:37:51.538800000
5 5 0.322344 12.0 1970-01-01 01:36:50.690822400
6 6 -0.670031 99.0 1970-01-01 13:33:29.236262400
7 7 0.551990 58.0 1970-01-01 06:13:30.637027200
8 8 -0.793331 86.0 1970-01-01 23:44:02.882140800
9 9 -0.009788 93.0 1970-01-01 17:43:11.392406400
... ... ... ... ...
490 490 -0.732420 97.0 1970-01-01 14:23:57.163315200
491 491 0.447224 55.0 1970-01-01 06:03:54.739900800
492 492 -0.481879 62.0 1970-01-01 07:47:52.262966400
493 493 -0.044106 99.0 1970-01-01 18:41:12.605395200
494 494 -0.989142 87.0 1970-01-01 20:09:18.645177600
495 495 0.499767 93.0 1970-01-01 22:41:12.925276800
496 496 -0.465699 101.0 1970-01-01 12:26:03.898233600
497 497 0.993191 59.0 1970-01-01 07:30:32.219683199
498 498 0.119690 92.0 1970-01-01 23:22:20.291030400
499 499 -0.127432 101.0 1970-01-01 16:31:42.817382400

The population table contains 4 columns. The column called column_01 contains a random numerical value. The next column, targets, is the one we want to predict in the analysis. To this end, we also need to use the information from the peripheral table.

The relationship between the population and peripheral table is established using the join_key and time_stamp columns: Join keys are used to connect one or more rows from one table with one or more rows from the other table. Time stamps are used to limit these joins by enforcing causality and thus ensuring that no data from the future is used during the training.

The peripheral table looks like this

  join_key column_01 time_stamp
0 26 -0.296267 1970-01-01 07:49:02.153308800
1 12 0.592168 1970-01-01 18:13:22.004803200
2 42 -0.985272 1970-01-01 08:40:46.879334400
3 295 0.226407 1970-01-01 04:33:29.892729599
4 321 -0.443054 1970-01-01 13:06:06.752217600
5 331 0.036371 1970-01-01 07:37:57.732297600
6 380 0.627330 1970-01-01 11:35:47.812905600
7 39 0.253938 1970-01-01 23:48:30.471840000
8 389 -0.952225 1970-01-01 00:56:13.548134400
9 281 -0.097608 1970-01-01 05:57:45.546307200
... ... ... ...
99990 362 0.435637 1970-01-01 17:47:24.640483200
99991 414 0.952651 1970-01-01 14:10:45.324163200
99992 123 0.602393 1970-01-01 06:09:30.276028800
99993 351 -0.234709 1970-01-01 03:31:29.799033600
99994 408 -0.828928 1970-01-01 10:40:38.872704000
99995 213 0.946612 1970-01-01 11:24:28.479772800
99996 16 -0.809395 1970-01-01 13:18:16.385011200
99997 102 -0.055429 1970-01-01 20:31:45.983539200
99998 482 -0.973701 1970-01-01 21:59:30.909408000
99999 84 -0.187830 1970-01-01 16:46:32.828534400

In the peripheral table, columns_01 also contains a random numerical value. The population table and the peripheral table have a one-to-many relationship via join_key. This means that one row in the population table is associated to many rows in the peripheral table. In order to use the information from the peripheral table, we need to merge the many rows corresponding to one entry in the population table into so-called features. This done using certain aggregations.

For example, such an aggregation could be the sum of all values in column_01. We could also apply a subcondition, like taking only values into account that fall into a certain time range with respect to the entry in the population table. In SQL code such a feature would look like this:

SELECT COUNT( * )
FROM POPULATION t1
LEFT JOIN PERIPHERAL t2
ON t1.join_key = t2.join_key
WHERE (
   ( t1.time_stamp - t2.time_stamp <= TIME_WINDOW )
) AND t2.time_stamp <= t1.time_stamp
GROUP BY t1.join_key,
     t1.time_stamp;

Unfortunately, neither the right aggregation nor the right subconditions are clear a priori. The feature that allows us to predict the target best could very well be e.g. the average of all values in column_01 that fall below a certain threshold, or something completely different. If you were to tackle this problem with classical machine learning tools, you would have to write many SQL features by hand and find the best ones in a trial-and-error-like fashion. At best, you could apply some domain knowledge that guides you towards the right direction. This approach, however, bears two major disadvantages that prevent you from finding the best-performing features.

  1. You might not have sufficient domain knowledge.
  2. You might not have sufficient resources for such a time-consuming, tedious, and error-prone process.

This is where getML comes in. It finds the correct features for you - automatically. You do not need to manually merge and aggregate tables in order to get started with a data science project. In addition, getML uses the derived features in a classical AutoML setting to easily make predictions with established and well-performing algorithms. This means getML provides an end-to-end solution starting from the relational data to a trained ML-model. How this is done via the getML Python API is demonstrated in the following.

Defining the data model

The next step is finding features in the data that allow an accurate prediction of the target variable in the population table. This is achieved using a MultirelModel. This model is responsible for the entire process from feature engineering to predicting the target variable based on the generated features. The MultirelModel requires a predefined data model in order to efficiently represent the data in memory. This is achieved via Placeholders. Placeholders are light and abstract representations of DataFrames and their relations amongst each other.

population_placeholder = population_table.to_placeholder()

peripheral_placeholder = peripheral_table.to_placeholder()

population_placeholder.join(peripheral_placeholder,
                            join_key="join_key",
                            time_stamp="time_stamp")

Now we can define the model. In addition to the Placeholders representing the DataFrames you also have to provide a predictor. Additionally, you can alter some hyperparameters like the number of features you want to train or the list of aggregations to select from when building features.

model = getml.models.MultirelModel(
    name='getting_started_model',
    population=population_placeholder,
    peripheral=[peripheral_placeholder],
    num_features=10,
    aggregation=[
        getml.models.aggregations.Count,
        getml.models.aggregations.Sum
    ],
    predictor=getml.predictors.LinearRegression(),
    seed=1706,
).send()

We have chosen a narrow search field in aggregation space by only letting the model use Count and Sum. For the sake of demonstration, we use a simple LinearRegression and construct only 10 different features. In real world projects you would construct at least ten times this number and get results significantly better than what we will achieve here.

Training a model

When fitting the model, we pass the handlers to the actual data residing in the getML engine - the DataFrames.

model = model.fit(
    population_table=population_table,
    peripheral_tables=[peripheral_table]
)
Loaded data. Features are now being trained...
Trained model.
Time taken: 0h:0m:0.157011

That’s it. The Multirel feature engineering routines as well as the LinearRegression contained in the MultirelModel are now trained on our test data set.

Scoring the model

Let’s generate another population table as validation data set in order to see how well the trained model performs on new data. For numerical predictions this results in three different scores: mean absolute error (MAE), root mean squared error (RMSE), and the square correlation coefficient (rsquared).

population_table_score, peripheral_table_score = getml.datasets.make_numerical(
    n_rows_population=200,
    n_rows_peripheral=8000,
    random_state=1710
)

scores = model.score(
    population_table=population_table_score,
    peripheral_tables=[peripheral_table_score]
)
print("Mean absolute error: {:.3f}".format(scores['mae'][0]))
Mean absolute error: 0.005

Our model is able to predict the target variable in the newly generated data set very accurately.

Making predictions

You can also make predictions using the model you have just trained

population_table_predict, peripheral_table_predict = getml.datasets.make_numerical(
    n_rows_population=200,
    n_rows_peripheral=8000,
    random_state=1711
)


yhat = model.predict(
    population_table=population_table_predict,
    peripheral_tables=[peripheral_table_predict]
)
print(yhat[:10])
[[ 5.00268213]
 [14.00858787]
 [24.00367308]
 [ 1.00441267]
 [27.00394183]
 [20.00157795]
 [16.00315811]
 [ 4.00264301]
 [19.00271721]
 [25.00384922]]

Extracting features

Of course you can also transform a specific data set into the corresponding features in order to insert them into another machine learning algorithm.

features = model.transform(
    population_table=population_table_predict,
    peripheral_tables=[peripheral_table_predict]
)
print(features)
[[ 5.          0.99524429  5.         ...  0.26285526  0.26285526
  -0.31856832]
 [14.          3.80100605 14.         ...  3.15211846  3.15211846
   0.39465668]
 [24.          5.29167009 24.         ...  5.92441112  5.92441112
   0.12470039]
 ...
 [ 8.          2.00532951  8.         ...  0.94089783  0.94089783
  -0.74996369]
 [15.          1.90051102 15.         ...  2.11328521  2.11562428
  -0.72788024]
 [ 2.          0.6167304   2.         ...  0.05360352  0.05360352
  -0.35370042]]

If you want to see the SQL code for each feature, you can do so by clicking on the feature in the monitor or calling the to_sql method on the MultirelModel. The definition of feature_2 is

CREATE TABLE FEATURE_2 AS
SELECT COUNT( * ) AS feature_2,
       t1.join_key,
       t1.time_stamp
FROM (
     SELECT *,
            ROW_NUMBER() OVER ( ORDER BY join_key, time_stamp ASC ) AS rownum
     FROM POPULATION
) t1
LEFT JOIN PERIPHERAL t2
ON t1.join_key = t2.join_key
WHERE (
   ( t1.time_stamp - t2.time_stamp <= 0.499323 )
) AND t2.time_stamp <= t1.time_stamp
GROUP BY t1.rownum,
         t1.join_key,
         t1.time_stamp;

This very much resembles the ad hoc definition we tried in the beginning. The correct aggregation to use on this data set is Count with the subcondition that only entries within a time window of 0.5 are considered. getML extracted this definition completely autonomously.

Next steps

This guide has shown you the very basics of getML. Starting with raw data you have completed a full project including feature engineering and linear regression using an automated end-to-end pipeline. The most tedious part of this process - finding the right aggregations and subconditions to construct a feature table from the relational data model - was also included in this pipeline.

But there’s more! Check out the documentation for detailed information on how to work with getML. There are also related articles showing application of getML on real world data sets.

Also, don’t hesitate to contact us with your feedback.