The deep learning revolution has enabled automated feature engineering for images and sound data. Yet, for relational data and classical time series analysis, feature engineering is still done by hand or using very simple brute force methods. Our mission is to change that.
The automation of feature engineering on relational data and time series is at the heart of the getML software suite. There are other libraries that implement feature engineering tools on top of frameworks like data.tables in R, pandas in Python, or Apache Spark. In essence, they all use a brute force approach: Generate a large number of features, then use some sort of feature selection routine to pick a small subselection of them.
getML has chosen another path: Our highly efficient feature learning algorithms produce features that are far more advanced than what manual feature engineering could achieve or what could be accomplished using simple brute force approaches.
Feature engineering is the process of construction variables, so-called features, from a dataset. These features are used as input for machine learning algorithms. In most real-world datasets, the raw data is spread over multiple tables and the task is to bring these tables together and construct features based on their relationships. These features are stored in a flat feature table. In other words, feature engineering is the operation of merging and aggregating a relational data model into a flat (feature) table.
This is sometimes referred to as data wrangling.
Traditionally, feature engineering is done manually, by using brute force approaches or domain knowledge. In any case it is a tedious, time-consuming, and error-prone process. Manual feature engineering is done by writing scripts in Python, R, or SQL.
getML’s approch to feature learning is fundamentally different. It provides a framework capable of automatically extracting useful and meaningful features from a relational data model by finding the best merge and aggregate operations.
Unfortunately, the term feature engineering is ambiguous. More often than not, feature engineering is meant to describe numerical transformations or encoding techniques on a single table. The definition used above assumes that the raw data comes in relational form, which is true for almost all real-world data sets.
The general procedure for feature learning on relational data and time series using getML looks like this:
The only required input is a relational data schema. In particular, there needs to be some sort of target variable(s), which shall be predicted. For time series, the schema would typically be a self-join. In addition to this general information on the data schema, the intended usage of the variables has to be provided by setting the roles of the corresponding columns. How to setup a data scheme is described in Data model.
Features are often of the type (illustrated in pseudo SQL-like syntax)
COUNT the number of `transactions` within the last X `days`
where is some sort of fixed numerical value. getML’s algorithms do identify appropriate values for automatically and there is no need for you to provide them by hand.
Features can also take the form of
COUNT the number of `transactions` for which the `transaction type` is ‘type_1’ OR ‘type_2’ OR ’type_3’ OR …
getML’s algorithms also find appropriate conditions based on categorical data without any input from the user.
The feature learning algorithms can handle combinations of conditions too. So, features of the form
SOME_AGGREGATION( over some column ) WHERE ( condition_1 AND condition_2 ) OR ( condition_3 AND condition_4 ) OR condition_5
will be engineered automatically as well. Again, no input from the user is required.
To increase transparency relating to the created features, they can be expressed in SQL code. Even though automatically generated features will always be less intuitive than hand-crafted ones and could be quite complex, we want the user to get an understanding of what is going on.
getML provides two powerful feature learning algorithms: Multirel and Relboost.
Simply speaking, Multirel is a more efficient variation of Multi-relational Decision Tree Learning (MRDTL). The core idea is to minimize redundancies in the original algorithm by incremental updates. We then combined our improved version of MRDTL with ensemble learning methods.
MRDTL is a strain of academic literature that was particularly popular in the early 2000s. It is based on a greedy, tree-like approach:
Define some sort of objective function that evaluates the quality of your feature as it relates to the target variable(s).
Pick an aggregation and some column to be aggregated.
Try out different conditions. Keep the one that generates the greatest improvement of your objective. Repeat until no improvement can be found or some sort of stopping criterion is reached.
The reason this approach has never really taken off outside of academia is that an efficient implementation is far from trivial. Most papers on MRDTL implement the algorithm on top of an existing relational database system, like MySQL.
The main problem with trying to implement something like this on top of an existing database is that it requires many redundant operations. Consider a feature like
COUNT the number of `transactions` in the last X `days`
As we iterate through different values for the threshold , we are forced to repeat the same operations on the same data over and over again. Tasks like this bring traditional database engine to their knees.
The core idea of getML’s Multirel algorithm is to minimize redundancies through incremental updates. To allow for incremental updates and maximal efficiency, we developed a database engine from scratch in C++. When we evaluate a feature like
COUNT the number of `transactions` in the last 90 `days`
COUNT the number of `transactions` in the last 91 `days`
very little changes in between. Multirel only recalculates what has changed and keeps everything else untouched. Therefore, it needs two ingredients that can be incrementally updated: An objective function and the actual aggregation(s).
Our first ingredient must be suited for incremental updates. When we move from 90 to 91 days, presumably only very few lines in the population table actually change. We do not need to recalculate the entire table. In practice, most commonly used objective functions are fine and this is not much of a limitation. However, there are some, like rank correlation, that cannot be used.
The second ingredient, the aggregations, must allow for incremental updates too. This part is a bit harder, so let us elaborate: Let’s say we have a match between the population table that contains our targets and another table (or a self-join). This match happens to be between the two thresholds 90 and 91 days. As we move from 90 to 91 days, we have to update our aggregation for that match. For maximum efficiency, this needs also to be done incrementally. That means we do not want to recalculate the entire aggregation for all matches that it aggregates - instead just for the one match in between the two thresholds.
We want, however, to also support the
combinations of conditions. So, it is possible that a match was not
included in the aggregation before, but becomes part of it as we move
the threshold. It is also possible that the match was included in
the aggregation, but now it isn’t anymore.
For an aggregation like
incremental updates are straight-forward. If the match was not
included, but now it is, then increment by 1. If was included, but it
isn’t anymore, then decrement by 1.
Things are more tricky for aggregations like
CountDistinct. For instance,
Max is easy,
decrementing it is hard. If the match used to be included and is in
fact the maximum value, we now have to find the next biggest
match. And we have to find it quickly - ideally iterating through a
set of thresholds should take linear time in the number of matches. To
make it even more complicated, some cross-joins might result in a lot
of matches, so any data structures that have non-trivial memory
overhead are a no-go.
Everything so far has shed light on how we train one feature. But in practice, we want more than one. So, how do we do that? Since we are using a tree-based algorithm anyway, we are able to harness the power of ensemble learning algorithms that have been shown to work very well with non-relational decision trees, namely bagging and gradient boosting.
With bagging, we just sample randomly from our population table. We train a feature on that sample and then pick a different random sample to train the next feature.
With gradient boosting, we calculate the pseudo-residuals of our previously trained features. We then train features that predict these pseudo-residuals. This procedure guarantees that new features are targeted and compensate the weaknesses of older ones.
Transpiled to SQL, a typical feature generated by Multirel looks like this:
CREATE TABLE FEATURE_1 AS SELECT COUNT( * ) AS feature_1, 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 <= 0.499624 ) ) AND t2.time_stamp <= t1.time_stamp GROUP BY t1.rownum, t1.join_key, t1.time_stamp;
Further information can be found in the API documentation for
Relboost is a generalization of the gradient boosting algorithm. More specifically, it generalizes the xgboost implementation to relational learning.
The main difference between Relboost and Multirel is that Multirel aggregates columns, whereas Relboost aggregates learnable weights.
Relboost addresses a problem with Multirel that is related to computational complexity theory: In Multirel, every column can be aggregated and/or used for generating a condition. That means that the number of possible features is in the number of columns in the original tables. As a result having twice as many columns will lead to a search space that is four times as large (in reality, it is a bit more complicated than that, but the basic point is true).
Any computer scientist or applied mathematician will tell you that is not nice. If you have tables with many columns, it might turn out to be a problem. Of course, this issue is not specific to Multirel: It is a very fundamental problem that you would also have, if you were to write your features by hand or use brute force.
Relboost offers a way out of this dilemma: Because Relboost aggregates learnable weights and columns will only be used for conditions, but not for aggregation. So, now the search space is in the number of columns in the original tables - much better.
This might seem very theoretical, but it has considerable implications: From our experience with real-world data in various projects, we know that Relboost usually outperforms Multirel in terms of predictive accuracy and training time.
However, these advantages come at a price: First, the features generated by Relboost are less intuitive. They are further away from what you might write by hand, even though they can still be expressed as SQL code. Second, it is more difficult to apply Relboost to multiple targets, because Relboost has to learn separate rules and weights for each target.
Expressed as SQL code, a typical feature generated by Relboost looks like this:
CREATE TABLE FEATURE_1 AS SELECT SUM( CASE WHEN ( t1.time_stamp - t2.time_stamp > 0.499624 ) THEN 0.0 WHEN ( t1.time_stamp - t2.time_stamp <= 0.499624 OR t1.time_stamp IS NULL OR t2.time_stamp IS NULL ) THEN 1.0 ELSE NULL END ) AS feature_1, 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 t2.time_stamp <= t1.time_stamp GROUP BY t1.rownum, t1.join_key, t1.time_stamp;
Further information can be found in the API documentation for
This is a listing of suggestions how to get the most out of getML’s feature engineering algorithms.
How to make features more interpretable¶
First of all, you should keep in mind that there is a trade-off between the interpretability and the predictive power of the generated features, just like there is a trade-off between interpretability and predictive power of machine learning algorithms.
MultirelModel offers numerous parameters to
regularize your features, to keep them readable, and to avoid
The most important ones are:
regularization: A higher value of this variable is the most
elegant and efficient way to regularize the feature engineering
algorithm and will lead to less complex features and less danger of
min_num_samples: Determines the minimum number of samples a
subcondition should apply to in order for it to be considered. Higher
values lead to less complex statements and less danger of
max_length: This parameter imposes a hard upper limit on
the length of your subconditions. Settings it to 1 or 2 will greatly
increase the readability of your features.
allow_sets: Multirel can combine categories into sets. This
will produce conditions of the form
(CATEGORY = "value_1" OR CATEGORY = "value_2" OR ...)
Enabling this option can be powerful, but it can also produce
features, which are hard to read. So, you might want to set
allow_sets to False if you strive for interpretability.
How to add handcrafted features¶
Which hyperparameters have the most impact for Multirel?¶
There are several things you have to keep in mind when working with
num_featuresis often more than you think.
From our experience with hand-crafted features we were used to have some 30 to 50 features. But many problems, particularly for business-related data sets, often contain millions of samples of labeled data. That means if we create some 1000 features, the ratio of features to samples will still be below 0.1%. So, the only thing that should stop us from generating 1000 features are memory limitations. But if you have sufficient memory, do try to generate as much features as you please and see if it improves your scores on the validation set.
share_aggregationscan make a difference.
Every time a new feature is generated, the aggregation will be taken from a random subsample of possible aggregations and values to be aggregated.
share_aggregationsdetermines the size of that subsample.
share_aggregationsis, thus, the parameter to control the accuracy vs. variance trade-off: A higher value results in more accurate features, whereas a lower ones in greater variance.
The ideal value can vary greatly with the data set. We have seen sets where 0.25 is appropriate, but we gained wonderful results with values around of 0.05 on others.
The most extreme way to get great variance at the expense of accuracy, is
round_robin: When it is set to True, you will be guaranteed to have a great variety of features (but many of them might be meaningless).
Put some effort into regularization.
You can regularize your features using
A good data model trumps everything else.
But the most important thing is to think about your data model. We have tried to make the Python API as easy and flexible as possible by introducing our Placeholder. Thinking about your data model (don’t forget about the units) is often a better use of your time than microtuning the hyperparameters. After all, you have got a hyperparameter optimization at hand that automates this task.
How to reduce the training time of Multirel¶
The following parameters can be used to reduce your training time:
share_aggregations: Every time a new feature is generated,
the particular aggregation will be taken from a random subsample of
possible aggregations and values to be aggregated. The parameter
determines the size of that subsample. A lower value will increase
training time significantly.
round_robin: Setting this parameter to True will almost
certainly reduce your training time.
max_length: These parameters are used to
features and keep them simple. Simpler features take less time to
train. So, anything you do to regularize them will also reduce
aggregations: Keep in mind that some
aggregations, such as
Median, can be far more expensive
than others. If you want to reduce training time, sticking to the “big
Max) might be a good idea.
grid_factor: A higher values will cause Multirel to try
more critical values for your numeric features, which can take quite
some time. But keep in mind that this time penalty is quite low
compared to the increased accuracy you get.
Influence of the data set on the training time¶
It may sound very counterintuitive, but size does not matter much! Multirel uses a bootstrapping approach and subsamples of your population table. A new subsample is generated for each feature. So, whether your population table consists of 500 or 50,000,000 lines does not matter much.
What, instead, does matter, is whether the number of unique join keys is small relative to the size of your data set. If there are a lot of entries to aggregate in your peripheral table for every entry in your population table (as you might get in a time series problem), that will take a while.
By contrast, if you are doing a churn use case with many customers and few activities per customer, that will be fast.