When Does Model Monitoring Take Place?

Machine learning (ML) models are increasingly being used for a variety of use cases across all industries. But all models will inevitably experience degradation and data drift over time, leading to underperformance. Model monitoring aims to alleviate that issue before it results in adverse impacts.

AI model monitoring alerts users when an issue in a model starts so that you can fix and retrain it if necessary. The model monitoring framework runs from the development phase all the way through the deployment, operation, and retraining phases so that any problems can be fixed quickly and your model can run as productively as possible.

Why is model monitoring needed after deploying the model into production?

Machine learning model monitoring in production as well as before deployment is needed due to model degradation. To continue to work effectively, your models must be monitored for any performance decline. Such issues may be caused by factors such as data drift and model bias.

Data drift (as well as model drift, feature drift, and concept drift) occurs when there is a significant change from training data to data ingested in production. These shifts happen because the assumptions used in the training stage no longer apply during production. They may be caused either by:‍

• ‍Changes in the system: Some systems change due to varying factors, such as the time of the year or even the economy. For example, models that predict customer buying behavior may inaccurately assume that you need more or less products than you do. If you trained your model in the summer, it may experience data drift when buying behavior significantly increases around the holidays. Or, if you trained your model when the economy was in good shape, it might experience drift if the economy takes a turn for the worse.‍

• Changes to the way the model interprets data: When a user changes something upstream, it can alter the assumptions that your model makes. For instance, if a user changes how the model tracks the amount of goods shipped from units of a particular item to pounds of that same item, the model will experience drift.‍

• Real-world data differs from training data: The data used to train a model does not always accurately represent real-world data. For example, you may train your model using data that suggests that customers prefer a particular service you offer over others. However, that data may only represent a small population and does not reflect that, overall, your customers prefer a different service.

• Data drift can lead to model bias, which occurs when your model makes an incorrect assumption. As more data enters your model, its algorithm may learn the wrong signals by considering only parts of the data. For example, a model that Amazon used to recruit employees was biased against women because most former applicants were men, and their resumes were used to train the model.

Model monitoring is necessary to stop data drift and model bias from occurring or to catch errors that the model makes so that you can retrain it for better, more accurate performance.

How do you monitor a ML model?

There are a few ways to monitor a machine learning model. For one, you can compare model outputs to ground truth. Ground truth refers to baseline data that can be used as the standard for the data your model outputs. For this method of monitoring, you merely have to calculate the accuracy of production data based on the standard you set. However, baseline data isn’t always available in the real world or may take too long to gather, so other methods of monitoring become necessary.

Another such method is to examine target variables and input features. That’s because a model degrades due to a changing relationship between input features and their target variables. To determine the accuracy of your model using model monitoring metrics, find the difference between your initial data set and a second one.

Models can also monitor by analyzing feature importance. Look at shifts in the importance of certain features as well as changes in the ways those features are ordered in importance.

It is important to note that on top of monitoring your output data, you should monitor your input data, too. The assumptions initially made when training your model may no longer apply based on what’s happening in the real world.

Why is ML model monitoring important?

Model monitoring of machine learning models is important for several reasons:

• It allows for faster time-to-market: With monitoring, you can catch and fix issues right away and better understand how to retrain your model.
• It reduces costs: When you monitor your ML models, you reduce the number of errors your model makes. That means less company time (hence, less money) that your team needs to spend fixing issues. It also means you can push out more models in less time, which allows you to grow your revenue.
• It eliminates data silos: When you use model monitoring tools like Fiddler, you will have all your processes and solutions in one platform. This will allow several people on your team to work on a single model at once and will alert your entire team in real-time when an issue arises.

To learn more about model monitoring and how it works, try Fiddler today.