Data Cleaning and Wrangling in SQL


By Antonio Emilio Badia, Assoc. Professor, CSE department, U. of Louisville.

Knowing SQL is considered one of the basic skills of a data scientist because a large amount of data exists (and continues to be collected) in relational databases. A typical approach to data analysis is to extract the data from such repositories and carry out the analysis in R or Python, using the powerful packages/libraries available in both environments. In such an approach, SQL is only used for the sub-setting and extraction of data; basic querying suffices for this.

However, practitioners know that there is a (long path from raw data to analysis: data must be carefully prepared, a complex task involving several processes usually including labeled data cleaningdata wrangling, or data pre-processing. A question arises as to in which environment to carry out this process. Some analysts prefer to carry out the whole process in R or Python, as they both provide abundant support for them and they are (especially R) highly interactive, while SQL lacks the rich packages/libraries that make many operations on data a simple one-liner in R or Python, and its syntax can be quite constraining. On the other hand, there are a few disadvantages to this approach: when dealing with very large datasets, both R and Python can be slow or run into problems, and the multiplicity of packages/libraries (many times with overlapping functionality) creates additional complexity in these environments. Therefore, in some cases, at least, doing some of the data cleaning and wrangling in SQL can be beneficial. As we show here, many common tasks can be accomplished without too much extra burden in SQL. To illustrate this, we provide an example that focuses on the identification and substitution of missing values.

Missing values are represented by the NULL marker in SQL, but data may not always be clearly marked. Imagine a dataset containing table Patients with information about patients in a medical study. One of the attributes is id, an identifier, and two others are Height and Weight, representing respectively the height and weight of each patient at the beginning of a study. We note that some weight values are missing, indicated by a -1. We then ‘clean’ the table as follows:

UPDATE Patients
SET Weight = NULL
WHERE Weight = -1;


 

Once this is done, we can use the predicate IS NULL to deal with all missing values uniformly: the query

 

SELECT count(*)
FROM Patient
WHERE Weight IS NULL;


 

will tell us exactly how many values are missing. If only a few values are missing, we may want to drop incomplete data. The SQL command

 

DELETE FROM Patient
WHERE Weight IS NULL;


 

would eliminate rows (observations) where the weight is missing. If the attribute is deemed not important for subsequent analysis, the command

 

UPDATE TABLE Patient
DROP ATTRIBUTE Weight;


 

would eliminate the attribute from the whole table–that is, from all rows (observations). If we prefer to impute the missing values, the command:

 

UPDATE TABLE Patient
SET Weight = (SELECT avg(Weight) FROM Patient)
WHERE Weight IS NULL;


 

would substitute all missing values by the mean of the existing values. A more sophisticated approach would be to impute the value of the missing attribute from a related one. We may wonder if Height is related to weight; the query:

 

SELECT cor(Weight, Height)
FROM Patient;


 

will compute the correlation coefficient between the two attributes in a system where a built-in correlation function cor exists. Even in systems where it doesn’t, SQL allows us to simply write our own:

 

SELECT (avg(Weight * Height) - (avg(Weight)* avg(Height))) / (std(Weight) * std(Height))
FROM Patient;


 

because the mean (avg) and standard deviation (std) functions are universally present. If we consider the correlation high enough, we can use the kNN algorithm to infer appropriate values. In a general case, SQL allows us to compute distances between the elements of a dataset using the usual distance functions (like Euclidean or another norm) and sort the result by the computed distances so that the k nearest neighbors can be used to extrapolate the desired value. In this particular example, the computation is very simple: we use the absolute value of the difference of heights as our distance, and the average weight of the closest 5 neighbors for our new, imputed weight:

 

SELECT id, avg(Weight)
FROM (SELECT P2.id, P.Weight
      FROM Patient P, Patient P2
      WHERE P2.Weight IS NULL and P.id <> P2.id
      ORDER BY abs(P.Height - P2.Height)
      LIMIT 5) as KNN;
GROUP BY id; 


 

In this query, the subquery in the FROM clause computes distances between two distinct data points (patients), with one of them having a missing weight; orders the result by this distance (ORDER BY) and keeps only the 5 closest results (LIMIT); the average of the weights of this 5 neighbors is then used for the imputed value (note: ties are broken arbitrarily). We return a result for each data point with a missing weight, so we provide the id of such a point. This result can be used in a second query, similar to the second UPDATE above, to change the data values in Patient.

SQL has the added advantage of being a well-established standard for operating rather efficiently on very large datasets (recall that a separate query optimizer analyzes all the statements above and determines the most efficient way that the system can execute the SQL command, given current data and resources). Being able to carry out some data cleaning, wrangling, and filtering before extracting the data from the database may make a data pipeline simpler and more efficient.

 

Bio: Antonio Badia is an Associate Professor in the Computer Science and Engineering department, University of Louisville. His research in databases has been supported by the NSF (including a CAREER Award) and resulted in over 50 publications. He teaches courses in databases and an introduction to data management and analysis for non-CS majors. He is the author of SQL for Data Science: Data Cleaning, Wrangling and Analytics with Relational Databases, Springer, from which this post is extracted.

Related:



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *