Let's board our Delorean and travel back in time a little. Not to 1955 or even 1985 but just a few months: April 2024. Back then, I started a series of articles aimed to explain the theory behind Qlik AutoML. I started explaining the explanation of Machine Learning predictions. In other words, my first article in this endeavor was about SHAP Statistics. A few months later, I wrote about Qlik AutoML Preprocessing tasks, and my very last article of 2024, I explored how a model performance is calculated. Now we're ready for the 4th article in this series. Prepare to learn more about this important concept as I drop a lot of 1990's references (tell me how many you got).
When a machine learning model is trained, tested, and validated, the next logical step is to deploy it so it becomes an integral part of your data-driven decision-making process. In a Qlik environment, it becomes an artifact you can call using data connections, automations, real-time inside an app or even externally via API.
People start using your model, and now your business dynamics shift: previously, decisions were made based on existing criteria and people's expertise, but this new model’s outputs become another key factor in how choices are made. It’s perfectly normal, even expected, that your model’s performance will begin to degrade over time, because the mere act of using the model changes the environment in which it operates. It is like when Marty McFly changes the future by simply traveling to the past.
This is when it’s time to retrain your model, now incorporating the effect of its own predictions into the training data. But how do we identify that our model’s predictive performance is declining in the first place? In other words, how do we spot when it’s drifting? Welcome to the world of data drifting monitoring.
What Is Data Drifting?
Data drifting (or simply “data drift”) is the subtle or not-so-subtle change in the properties of incoming data compared to the data that was used to train your model. If, in 1990, you built a recommendation engine for video rental stores, you trained it on the tastes and preferences of people renting VHS tapes. Now it’s 2025, streaming is king, and VHS is ancient history. Even if people still occasionally dust off a VCR for “Back to the Future,” the model trained on VHS-era preferences will be out of tune with how people consume media today. That mismatch is data drift at work. Data drifting is important for a few reasons.
Decreasing Accuracy: as new data diverges from the training data, your model’s predictions can become less reliable. It’s like trying to navigate modern roads with a map from 30 years ago: some of the streets may still be the same, but a lot has changed.
Increased Bias: if the model is never updated to reflect changes, new biases could emerge, or old biases might get amplified. This can result in skewed decisions that no longer align with reality.
Cost and Maintenance: ignoring drift often leads to an expensive and time-consuming overhaul rather than routine maintenance. Regular monitoring and timely retraining help keep these costs manageable over the long haul.
To make it absolutely clear: data drifting is not necessarily bad. Changes are part of the process. Maybe it happened simply because the world, business, products and clients are always changing, or it could simply be a sign that the model is being adopted. In that case, change is exactly the point. What is bad is not monitoring or not acting upon change.
Data Drift vs. Concept Drift
In regard to changes in data over time, there are actually two different phenomena that can mess with your models, like how people often confused Nirvana's "Smells Like Teen Spirit" with Pearl Jam's "Even Flow" back in the day (even though true rock and roll fans knew the difference immediately), or when people thought the Guns n' Roses wrote "Live or Let Die" (this one is particularly annoying to me).
Data drift happens when your input data changes but the relationships stay the same. It's like how in Back to the Future, Marty McFly lands in 1955 where the town looks different, the cars are different, and people talk differently, but the fundamental rules of how Hill Valley works haven't changed. The features of your data have shifted, but what they mean in relation to each other remains consistent.
Concept drift, on the other hand, is when the relationships between your variables change. This would be like if Marty landed in an alternate 1985 where suddenly Mr. Strickland was super chill and giving kids high-fives instead of detention. The underlying concept of how things relate to each other has fundamentally changed - and that requires a whole new approach. Imagine a new product or unit is launched or closed. This might change the whole relationship and dynamics in the data.
Concept drift can sneak into your models in a few different styles, each with its own personality. Sudden drift is like flipping a switch: one day your data behaves normally, the next it’s completely different (think of a market crash or policy change). Gradual drift is more subtle, where the data slowly changes over time, almost like background music getting louder until you realize it's a whole new vibe. Incremental drift is a cousin of gradual drift but more stepwise. Data shifts in small, consistent chunks, like leveling up in a video game. And then there’s recurring concepts, where old patterns come back around like flannel shirts or vinyl records. Your data reverts to previously seen behaviors, which can trip up models that assume everything always moves forward.
Early research on these concepts dates back to the 1980s and were advanced in the 1990s, coincidentally enough (and that's why all of my references), right when Wayne and Garth were partying on, Bill and Ted were having excellent adventures, and researchers like Schlimmer and Granger and Widmer and Kubat were pioneering work on adaptive learning systems. Pretty fly for computer science guys!
Why Data Drift Happens
Data drift occurs for all sorts of reasons: many of them simply reflecting a world in constant change. Understanding these causes is essential if you want to keep your models relevant and accurate over the long haul.
Changing User Behavior
People’s behaviors can shift surprisingly fast, much like the way 90s pop culture pivoted from Vanilla Ice to Nirvana almost overnight. An e-commerce recommendation engine trained on data from the Tamagotchi-obsessed era won’t stay current if customers have moved on to entirely new trends.
Seasonal Patterns
Many data patterns follow predictable cycles: retail spikes around holidays, energy consumption varies by season, and commuter traffic changes between weekdays and weekends. Like the steady drumbeat in a classic Metallica track, these seasonal shifts are expected (Sorry Lars, it kind is predictable), but you still need to adjust your model to stay in sync.
Evolving Data Collection Methods
Sometimes the world isn’t what’s changing but your measurement approach is. Giving more attention to patients predicted to get an infection, altering survey questions, or changing HR procedures can cause data distributions to shift. It’s a bit like thinking "if Nirvana never existed, would Green Day be different and therefore Emo would never exist?".
When Data Drift Undermines Your Models
When data drift goes unnoticed, even a solid machine learning model can gradually become less effective, impacting critical decisions across various industries.
Diminished Accuracy
As drift progresses, your model’s accuracy erodes, similar to how MTV shifted from music videos to reality TV, so gradually that viewers hardly noticed until everything had changed to the point that the name MTV does not mean anything anymore.
Real-World Consequences
A risk model might offer loans to the wrong applicants, healthcare models could make faulty diagnoses, and recommendation engines may suggest products that users have zero interest in. It could also create a skeptical attitude towards all the Data Science and IA initiatives within the organization.
Worsening Over Time
A small divergence between training data and production data can balloon into a major issue if left unchecked. Just as VHS tapes were once cutting-edge but soon became obsolete, an unchecked model can become unusable faster than you’d expect.
Detecting Data Drift
Spotting data drift is critical for maintaining reliable predictive systems.
Statistical Analyses
Compare current data properties (means, variances, distribution shapes) to those of the reference or training set. Significant deviations indicate your data may have changed enough to warrant action.
Domain Classifiers
Train a model to distinguish between your “old” reference data and newer production data. If it can consistently tell them apart, that’s a clear sign your data has drifted.
Monitoring Model Performance
Keeping track of key metrics (accuracy, precision, recall) over time can reveal a gradual slide in effectiveness. Much like how MTV’s music videos drifted from Nirvana’s grunge vibe to Britney’s pop songs.
Responding to Data Drift
Once you confirm that drift is happening, the next step is to update your modeling strategy.
Regular Model Retraining
Periodically refresh your model with current data to keep pace with new trends. Think of it as evolving your sound over time, much like Metallica did throughout the 90s. They went from raw thrash metal roots in the first albums to a more polished, mainstream kind-a-pop style, yet managed to hold onto a loyal following despite shifting their creative direction.
Adaptive Learning
Certain algorithms are designed to learn incrementally, adjusting continuously without the need for a full retraining cycle. It’s reminiscent of how the T-1000 in “Terminator 2” adapted to new threats on the fly. Humm, since we're discussing AI, maybe this was not the best example.
Sliding Window Approaches
Use only the most recent data points to train your model, systematically dropping outdated information. It’s akin to how long-running TV shows phased out early 90s fashion for late 90s styles, always staying somewhat up-to-date.
Ensembles
Combine multiple models trained on data from different time periods. By giving more weight to the models performing best on today’s data, you can continuously adapt, somewhat like how Josh Baskin in Big blended his boyish enthusiasm with his new adult responsibilities, ultimately finding a winning approach that balanced both worlds (great movie, by the way, but I am not sure how we were allowed to watch it when we were kids).
How Qlik AutoML performs Data Drifting
So how does Qlik AutoML pull a MacGyver and handle drift without bringing down the whole predictive party?
Automatic Monitoring (Because Manual is So Last Century)
Once you deploy a model in Qlik AutoML, it automatically begins monitoring for drift with no Tamagotchi-level babysitting required. It evaluates drift in input features, target variables, and even your model’s prediction confidence over time.
Drift Scores (Cue the MTV Chart Countdown)
In Qlik AutoML, data drift is measured using something called the Population Stability Index (PSI). Wait, what is PSI? No worries, Marty, let's understand it in plain English.
Imagine you trained your model during a glorious summer of 1998, when everyone was using Yahoo Mail and drinking Sprite. Your model learned patterns based on that world. Now it’s 2025, and your users are all on TikTok and mocha pumpkin lattes. PSI is a number that compares the distribution of a feature (like “Age” or “Product Category”) in your training data vs. your current data. It's basically asking:
“Are the values I’m seeing now way different from what I was trained on?”
The formula bins your data into ranges and looks at the percentage of values in each bin, then compares old vs. new.
Now for the part that’s as important as rewinding your VHS before returning it:
PSI < 0.1 → You’re good. The feature distribution hasn’t changed much.
0.1 ≤ PSI < 0.25 → Mild drift. t’s not panic time, but it’s worth watching.
PSI ≥ 0.25 → Whoa, Nelly! This feature has changed a lot. drifted enough that your model might be out of touch. Qlik AutoML suggests you retrain your model or spin up a new experiment.
Visualization Tools
You get a clean dashboard where Qlik shows the trend of each feature's drift over time, like watching your pager light up with alerts. It even highlights which features have drifted the most so you can act fast.
Actionable Insights
Unlike Clippy, Qlik AutoML doesn't just say “It looks like your data has drifted!” It suggests retraining the model when drift becomes significant, especially for high-importance features. It also logs every drift event, which is more documentation than most of us kept in high school.
TL;DR: Keeping Your Model from Going Full VHS
Machine learning models don’t age like fine wine, they drift. Data drift is what happens when the data your model sees in production starts acting differently from what it saw during training. This can tank your accuracy, increase bias, and turn your high-performance model into a confused relic. Concept drift kicks it up a notch by changing the actual relationships between variables. Whether the drift is sudden, gradual, incremental, or takes the form of recurring patterns (like that 90s band that just won’t quit), you need to monitor it. By the way, this applies to any relationships, especially marriage. it will drift eventually, monitor it and revive it. ;)
Qlik AutoML makes this easy (the data part not the marriage) by automatically tracking drift using PSI (Population Stability Index). PSI < 0.1? Chill. PSI between 0.1 and 0.25? Keep an eye on it. PSI ≥ 0.25? Time to retrain, friend. With automated monitoring, clear visualizations, and actionable alerts, Qlik AutoML keeps your model from becoming the predictive analytics equivalent of a beeper. Stay sharp, retrain when needed, and your model will keep delivering, no mixtape required.
So, whether your data is shifting like fashion trends or flipping like a surprise plot twist, staying on top of drift ensures your models keep making decisions that actually make sense, because in the world of machine learning, relevance isn’t forever, it’s monitored.