- Why Do ML Models Fail After Production?

In today’s transforming digital world, a wide variety of businesses leverage AI to perform forecasts on their sales, inventory, customer demand fluctuations, or maybe to provide recommendations to users while they shop.

Consider any particular business scenario, where you have been tasked to build and deploy a machine learning model. What comes to your mind when we think of any ML project pipeline?

Data collection & Processing -> EDA -> Feature selection/manipulation -> Model training & Tuning ->Model deployed to production

However, a very crucial step is skipped in the above pipeline, which is “Monitoring post-production”. In most of the cases, the failure of ML models after deployment can be attributed to either machine failures or degradation in model performance. While machine failures occur due overheating or hardware shutdown, degradation in model performance is trickier and can happen slowly over multiple years as the data distribution shifts over time. A model which had accurate and reliable performance during the time of deployment can easily become unusable if left unchecked and not re-trained periodically.

- Data Distribution Drifts : Why Should You Handle Them?

When we train a model on our curated training dataset, the model learns the patterns and the distribution of the underlying data. We also ensure that the model is able to use these learnings accurately on the test data distribution. But, when the model is already in production, it comes across real world “unseen” data (your wild wild west kind) and this real world data is not stationary by any means, i.e, its distribution is continuously shifting. This shift is termed as “ Data distribution drifts”. These shifts occur due to a variety of factors such as the change in everyday dynamics of the global economy, varying social norms and cultures across the world.

Let’s look at a quick example. Prior to 2022, when people searched for “ Elon Musk twitter”, they probably were looking for the trending tweets posted by Musk, or his twitter profile. But now, people are looking for information on Elon musk’s takeover and organizational changes at Twitter. This is an example of a shift in the data distribution. A key thing to note is that shifts in data do not occur just due to sudden events, they also occur in a seasonal and gradual way. The data underlying the pricing of retail properties in semi-urban areas would shift gradually. The behavior of your users will evolve over time as you make product innovations, needs of your users evolve over time due to cultural shifts, your competitors change their pricing strategies etc. If you don’t handle these drifts on time, the model would degrade and may create more loss than value.

- Types Of Drifts

The data distribution shifts we come across can be broadly grouped into Concept drift and covariate drift.

Concept drift : It occurs when the relationship between the inputs and target variables changes over time. Consider a model built to predict the rent of office spaces in San Francisco based on features like location, square footage, furnishing, etc. As companies are embracing the remote work culture, work from home drastically reduced the need for office buildings and the rents (target) went down. Even though the input features distribution remains the same, the model would perform poorly. This is an example of concept drift.

Data drift: Also known as covariate drift, it occurs due to a change in the underlying distribution of your input variables. Consider a case where you are building a model to predict if a new user on your website will be converted to a paid user or not. The model is built upon user features such as age, location, income, marketing channel, product interactions etc. Suppose you launch a marketing campaign which has attracted a lot of users with lower income compared to your current user-base. Your model is expected to underperform for these new set of users as it hasn’t the relationship between the input features and purchasing decision for this new set. This is an example of covariate drift

- Statistical Methods To Detect Data Drifts

As we now know the significance of handling data distribution drifts, the first step of the process is to identify and detect these shifts. The simplest way of detection is to monitor statistical values like mean, median. But these methods are insufficient for most of the ML use-cases. However, there are many statistical methods, more suited for changes in data distributions which can be used

a) Population Stability Index(PSI): It is a metric that measures the shift in population over time between the target (test data) at the deployment time and the target at the current period. The below image shows an example of how the population distribution has changed.

How do we quantify this?

We split the population distribution into buckets ( 10-20 based on the range) . For each bucket, we compute the percentage of old and new data population. Next, we apply the below formula for PSI calculation on each bucket, and sum their values for the final PSI metric.

If the above calculated PSI value is less than 0.1, the drift in population is negligible. When the value lies between 0.1-0.2, the drift is moderate. When PSI is higher than 0.2, there is significant drift in the population and we need to handle it. This metric for drift identification is very popular in financial data based ML models.

b) Kullback–Leibler Divergence: The KL divergence is a statistical metric, which measures the difference between probability distribution of old data( testing time) and new data (post-deployment in production). It is represented by “||’. This metric is often referred to as ‘relative entropy’ , as it uses the entropy of probability distribution in its calculation.

Recall that, for a discrete probability distribution, the entropy is given by the blow formula:

This expression is extended to compare the probability distributions of old data (Q) and new data (P) as shown below.

Note that KL divergence is asymmetric, you would end up with different results if you swap P and Q.

The value of the metric can range from 0 (ideal-case) to infinity. The larger the value, the more is the drift in probability distributions between both datasets.

c) Jensen-Shannon Distance: The JSD also measures the difference in probability distributions similar to KL divergence, with two major changes. The JSD metric calculates a normalized score using the KL divergence formula discussed in the previous section, as shown below.

Due to the normalizing, the value of the metric is restricted from 0 to 1, which makes interpretation easier for us. Also, this metric is symmetric adding another advantage. When the value is zero, it indicates the distributions are identical and there is no drift. The value of 1 corresponds to maximum drift.

d) Kolmogorov-Smirnov Test: The K-S test is different from other tests as it is a non-parametric test. What do we mean by non-parametric tests? They do not make any assumptions on the nature of underlying data distribution.

The Kolmogorov-Smirnov test measures the difference/distance between the cumulative probability distributions of the old and new data.

As it compares distribution of two data samples and performs null-hypothesis testing it is referred to as ‘two-sample test’. This method is best for implementing numerical data. A limitation of this method is its sensitivity in large datasets.

e) Page-Hinkley Method It is a commonly used method to detect drift in incoming data and provide warnings to the ML teams. In this method, the mean value is dynamically updated as there is an inflow of new data. At any point when the mean value would exceed the threshold value ( set by the user based on mean observed during deployment), it provides drift detection alarms. In practical usage, this method leverages the CUMSUM control chart to identify shifts.

To implement this in python, we can use the river library’s modules. It provides an easy way to control parameters such as no of instances to consider, the threshold value and so on.

- Conclusion : When To Use What?

Apart from knowing the implementation of these methods, it is essential to understand the best statistical test for different scenarios. The KL divergence does not depend upon the sample size, and is a good fit for large datasets. The Jenson-Shannon distance is more sensitive than the former , but still stable, reaching a good compromise. It is preferred to use the Kolmogorov-Smirnov test when we have a small sample size and we want to spot sensitive changes. Wasserstein distance is another statistical measure, which is useful to detect drift in higher dimensional spaces like images. Apart from these, there are model-based and time-window distribution based approaches for drift detection. Hope you enjoyed the read!

The above mentioned are some of the most commonly used statistical measures that work for more than 90% of ML models. UpTrain is an open-source toolkit which has out-of-the-box implementations for all these measures and provides a single line interface for you to start monitoring drifts in your ML models. Try it out here