Author: Abhishek Sengupta, Walmart Global Tech India
A necessary but often tedious task data scientists must perform as part of any project is quality assurance (QA) of their modules so as to prevent any unforeseen incidents during deployment. While Quality Assurance and Excellence is quite prevalent in Data Engineering, QE in Data Science world is still in an evolving state and is not always automated. For many reasons, it is neither possible nor desirable to completely automate QA. For example, thorough QA typically involves leveraging a degree of domain expertise which is not easily encoded in a computer program. However, we can still be freed of some of the tedium of QA. In this article, we illustrate a way to mix automation with interactivity to reduce the manual steps in QA without compromising the data scientist’s ability to do that which is best done manually.
Quality in Data Science
We break down this article into the below themes:
Starting off with the Data Science Architecture, the below diagram shows a generic Data Science Flow and its key components:
Each of these phases produce certain Artifacts as illustrated above. Our goal is to ensure and enhance quality of each of these artifacts by:
Before diving into the QC Part, let us first familiarize ourselves with 6 key dimensions of data quality which are applicable for most data sources:
Having understood the various dimensions for data quality, the idea is to pass all data tables through tests which provide a score on each of these data dimensions, with additional tests augmenting these tests to provide more nuanced reports. Each of these tests will have a predefined set of thresholds, and if the quality falls below a certain threshold, automated warnings will be triggered.
Let’s illustrate the above with an example. Consider the below table containing the sales and starting inventory for an item for the latest week:
Though there are multiple libraries which perform automated DQ checks, we lean towards PyDeequ , a Python API for Deequ, a library built on top of Apache Spark, mainly for its ability to handle scale and ease of integration into PySpark. PyDeequ provides a wide variety of Analyzers (https://pydeequ.readthedocs.io/en/latest/pydeequ.html#module-pydeequ.analyzers), which computes metrics for data profiling and validation at scale.
Below is a sample DQ Check on some of the columns for the above dataset using PySpark:
The code snippet tests for the below:
The result is as follows:
Note that PyDeequ does not have Analyzers specific to Accuracy and Timeliness, but user-defined compliance Analyzers can be used to serve the same purpose.
The next step is to set warning thresholds for each of these tests (each row), with alerts being trigged if the “value” falls below the set threshold.
Apart from these metrics, additional QA checks can also be performed based on additional constraints such as:
These tests can be performed similarly using the Compliance Analyzer from PyDeequ.
The other type of data DQ Checks that can be performed are Model Data Checks, which are specific to datasets which act as outputs of a Data Science Model. These can be categorized into the two below buckets:
Let’s illustrate this with another dataset, which contains the medical insurance “charges” for a set of persons with differing features. The goal is to predict the “charges” for a new person with known features. Below is a sample from the data:
For performing the model DQ Checks, we will take the help of PyTest, a Python Library for writing small, readable tests, which can scale to support complex functional testing for applications and libraries. Note that this can also be used in PySpark via Pandas UDFs.
The above example shows pre_train tests for data shape and leakage checks using PyTest . We have excluded user-defined functions like data_preparation, etc. for the sake of brevity. Also note that, instead of assert, we have pytest_check. It acts as a soft assert meaning that even if an assertion fails in between, the test will not stop there but instead proceed till completion and report all the assertion fails.
Post-train tests can be similarly defined using PyTest once the possible tests are identified. For our example, we shall check for sex invariance in the post-train test script. Ideally, the charges should not vary according to the gender of the individual applying for insurance. Let’s test for it by writing this test script.
In the above test script, we define two data points one with sex as “female” and the other with sex as “male” with all other values being equal. Now, we predict the charges using the model and check whether the charges are equal or not. Similar tests can be conducted for other features as well.
Data Fairness is also a big concern in Data Science which ensures that we use data in a way that doesn’t create or reinforce bias. Our society is full of biases, and some can be positive and improve the way we act towards others. Some biases are often unconscious and can often be replicated to software or program. Below are the broad categories of biases a modelling data can fall prey to:
What-if Tool is a great option to detect such biases through UI and threshold based options. Let’s take an example of income prediction, wherein a logistic regression model is trained on both sexes to predict their respective income as High or Low. Choosing the cut-off as 0.5 (>=0.5 – High Income) for both males and females result in 23.6% of males and only 8.1% of females getting classified as high income(As shown below in the left dashboard). This is a classic example of sampling bias wherein the large imbalance between working men and working women in the training data have resulted in men much more likely to be labelled as high income than women.
Using the What-if Tool, we can experiment with different cut-offs to arrive the optimal bias-free threshold. In our example, adjusting the thresholds (0.74 for men and 0.32 for women) results in a much more even distribution (12.1% for men and 11.8% for women) as can be seen on the right dashboard.
All such slice-and-dice on the data and parameters can be performed through sliders, clicks and drop-downs in the What-if Tool to help identify and alleviate these biases and ensure a fair modelling data.
In this article, we talked about the different dimensions of Data Quality and process to automate the same from a Data Science Lens. A robust QC Process for Data Science will go a long way to ensure quality releases, reduce bugs and improve trust and confidence in the solution. Automation of the same would reduce the time spent in ensuring that the solution adheres to the Quality Standards.
Title picture: freepik.com
3AI is India’s largest platform for AI & Analytics leaders, professionals & aspirants and a confluence of leading and marquee AI & Analytics leaders, experts, influencers & practitioners on one platform.
3AI platform enables leaders to engage with students and working professionals with 1:1 mentorship for competency augmentation and career enhancement opportunities through guided learning, contextualized interventions, focused knowledge sessions & conclaves, internship & placement assistance in AI & Analytics sphere.
3AI works closely with several academic institutions, enterprises, learning academies, startups, industry consortia to accelerate the growth of AI & Analytics industry and provide comprehensive suite of engage, learn & scale engagements and interventions to our members. 3AI platform have 16000+ active members from students & working professionals community, 500+ AI & Analytics thought leaders & mentors and an active outreach & engagement with 430+ enterprises & 125+ academic institutions.