Data drift is the concept of monitoring changes to your data over time. Data drift happens when the attributes of the data used to train your model change. This can happen for a number of reasons and is completely normal, but it’s important that you understand how it works so you can make sure your data will continue to be representative of what you’re trying to measure.
Data drift is a natural part of any analytics or machine learning solution. It happens because the world changes, and the environment around your data changes with it. In order to keep up, you need to make sure that your training data accurately reflects these changes in order to maintain accurate results.
Data drift monitoring is a process of tracking data changes over time to ensure that models are performing well. If you’re not monitoring for data drift, it can be difficult to detect, resulting in poor performance and inaccurate predictions.
Data Drift and Concept Drift
Data drift, also known as concept drift, is the phenomenon where the statistical properties of a given data set to change over time. This is in contrast to the more commonly recognized notion that entities or objects within a data set may change over time.
While most data science techniques rely on stable and predictable statistical properties, many real-world data sets are subject to data drift. This can occur over time, but it may also shift rapidly. Hence, the term drift is used to describe these variations in statistical properties.
While we have an intuitive understanding of how objects within a data set can change over time - for example when someone changes jobs or moves house - the more challenging question is how we can identify when and where a data set drifts statistically?
What Causes Data Drift?
There are many reasons why your data may change over time. The most common are:
Missing, corrupted, or bad data: Sometimes your training data can become corrupted or simply bad, which will affect the accuracy of your model.
Business process change: Business processes change all the time, especially if they’re manual and rely on human operators. Processes may have changed significantly since the last time you trained your model and not updated it accordingly.
External factors: There are many external factors that can influence how your data behaves and changes over time, some of which you may not even be aware of beforehand.
Data drift can happen at any stage of the data lifecycle:
Data storage: Data drift occurs when the nature or format of your data changes as it moves through various stages of the data lifecycle. For example, if you used a Python script to extract data from a database, then stored the output as JSON, it would be important to monitor any changes that might occur over time.
Data ingestion: During ingestion, new records appear in your dataset and existing ones are updated or removed. In this case, it’s important to make sure that you have enough historical data for training purposes and that these records continue to be accurate.
Why is Data Drift Important?
We use machine learning algorithms to learn patterns from our data. These patterns may be used for predictive purposes - inferring unknown values from known ones; descriptive purposes - identifying subgroups with similar characteristics; or prescriptive ones - recommending items based on past choices. If a model has learned patterns from data collected at one point in time that are no longer representative of current conditions, then it will produce poor results when applied to new situations.
About the Author -
Qualdo™ helps you to monitor machine learning models & data issues, errors, and quality in your favorite modern database management tools.
Top comments (1)
Good article emphasizing the importance of Data drift.