Quality data beats quality algorithms – this simple fact underscores the problem all data scientists know too well. Without high-quality data, their models are effectively useless.
Unfortunately, their colleagues who are tasked with sourcing the data don’t have too many options, and if you asked many of them off the record, they would gladly switch roles with Sisyphus, the figure from Greek mythology, doomed to eternally roll a boulder up a hill only to have it to roll down every time it neared the top.
Legacy suppliers, public resources, and many others approach data with the same mentality, “do you want it, or not?” leaving data scientists with the messy, time-consuming task of cleaning heaps of historical data to uncover the representative sample they need.
Fifty years ago, it would’ve been impossible to predict the value data represents today. Why would anyone have been concerned with database architecture when the societal expectation for progress was the development of flying cars and the establishment of space colonies?
While we aren’t there yet, you’d be hard-pressed to find a company today that would admit to not being data-driven. However, we have amassed 2.7 zettabytes – one sextillion or 1,000,000,000,000,000,000,000 bytes of data in our digital universe most of which is garbage.
A survey of data scientists by machine learning and AI company Figure Eight (formerly known as CrowdFlower) revealed that approximately 80% of a data scientist’s time is spent simply collecting, cleaning, and organizing data. Only 20% of their time is spent on more creative activities like mining data for patterns, refining algorithms, and building training sets.
In other words, for each day spent analyzing data, data scientists spend four days doing data preparation.
FigureEight (CrowdFlower) Data Science Report
This digital janitor work is by far the most time-consuming part of a data scientist’s job. While it’s certainly an important aspect of the job, you can’t trust the outputs of messy data hence the cleansing process. However, This process prevents data scientists from spending more of their time on higher-value activities that have a direct impact on their company’s operations and strategy.
Therein lies the challenge: How do we decrease the amount of time spent on data wrangling in order to increase the amount of time spent on value-generating activities?
The solution is schema standardization enforced by a third party.
Although data is widely considered to be highly complex by the general population, those in the industry know data ultimately consists of two components: names and types.
- Names are the different attributes of each data point. For example, if we’re looking at a data set from a vehicle information portal, it will have the vehicle information number (VIN), name, make/model, type, color, etc. These names organize the data and provide access to it for analysis.
- Types describe the attributes of each data point. In some instances, types have subtypes like units such as miles per hour versus kilometers per hour. Types can include text, such as a string like the VIN e.g. 2GCEK13T961100610, or a boolean, true or false. If the attribute is a number, it could be an integer, 10, or a floating-point number, 3.14159. It could also be a selection from an enumeration, which is a static set of values like color: red, yellow or blue. These types describe the names and dictate how a data scientist interacts with the attributes of a data set.
Names and types comprise the schema of any given data set. When data scientists build models for analysis, they are often incorporating multiple raw data sets from various sources. Each of these data sources uses their own schema, which results in a vast array of incompatible names and types. Before they can get to the fun part, data scientists must standardize these attributes across each source to produce a representative sample or compatible data set. On the one hand, this process prevents garbage in, garbage out, and on the other, unfortunately, it absorbs 80% of the workweek for data scientists.
When data sources adhere to schema standardization, however, names and types are innately compatible meaning all data sets work together from the beginning. Data from multiple sources can be normalized into a single feed and record level provenance by source can now be retained. Rather than spending the bulk of their time collecting and cleaning, data scientists can skip this step and go to data transformation, mining, and/or modeling. Schema standardization eliminates the time needed for data scientists to reformat and match the data so a majority of their time can now be spent on value-generating activities.
By leveraging a third-party offering that enforces schema standardization, and the ability to transform data in-house, data scientists can win back hundreds of hours of productivity, reduce costs associated with errors, and most importantly, help their organization make well-informed decisions.
About the Author:
By Nathan Dolan, Technical, Partner Success Associate, Narrative