With the exponential growth of data available on the web, user interests are constantly evolving. From discovering new hobbies to seasonal shopping patterns, web personalisation systems need to adapt quickly — without losing track of what users loved yesterday.
However, most machine learning models, especially neural networks, tend to forget what they learned before when introduced to new data. This phenomenon is known as catastrophic forgetting, and it’s one of the biggest challenges in continual learning.
That’s where Elastic Weight Consolidation (EWC) steps in — a game-changing technique that helps models remember the past while learning the future. And it’s especially powerful in web personalisation.
🤖 What Is Catastrophic Forgetting?
Elastic Weight Consolidation (EWC) is a technique that allows models to learn new tasks while preserving critical knowledge from previous ones.
Let’s say your AI model has learned that a user loves fitness gear — yoga mats, resistance bands, protein shakes. But a few months later, the user starts browsing kitchen appliances. If the model only trains on the new kitchen data, it might completely forget the fitness preference and stop recommending those items.
This is catastrophic forgetting: when a model overwrites old knowledge in favour of new learning, resulting in a drop in personalisation quality.
🧮 The Basic Idea
When a model finishes learning Task A (e.g., recommending fitness products), EWC computes how important each weight was for that task using the Fisher Information Matrix.
Later, when learning Task B (e.g., kitchen gadgets), EWC:
- Adds a penalty to the loss function
- Prevents changes to weights that were crucial for Task A
- Allows other weights to adapt freely
This balance keeps the model elastic — it can stretch to learn new tasks, but not in ways that damage previous knowledge.
🌐 Why EWC Is Critical in Web Personalisation
In web personalisation, user interests shift over time, often in non-linear and unpredictable ways. Here’s how EWC can make a huge difference:
- 🔁 Remembering Long-Term Preferences
Users don’t always stick to their latest obsession. A customer who’s into gardening today may return to fitness or fashion tomorrow. EWC ensures your model keeps long-term preferences in memory, avoiding one-dimensional recommendations.
Without EWC:
Only the most recent behavior is used — old preferences are forgotten.
With EWC:
The model adapts to new interests but keeps old ones accessible and useful.
- 🛒 Boosting Cross-Domain Recommendations
If your platform sells everything from books to electronics, EWC helps create smarter cross-domain recommendations. A user interested in productivity books might later browse smart home gadgets — and EWC ensures both insights are retained for richer, more varied suggestions.
- 🧍 Personalised Journeys That Grow with the User
User journeys are personal and cumulative. EWC enables personalisation systems to build a memory bank of what each user has explored — turning personalisation into a long-term relationship, not just a snapshot.
Imagine a user who:
- Buys a yoga mat in January
- Orders an air fryer in March
- Browses camping gear in June
With EWC, your system doesn’t forget the yoga mat by the time summer arrives — it might even suggest healthy camping snacks based on both fitness and cooking interests.
- 📉 Less Reliance on Data Storage
Storing all historical user data for retraining can be expensive, risky (privacy), and inefficient. EWC reduces this need by preserving knowledge in the model’s weights, not just in data.
That makes it a lightweight solution for personalisation at scale.
⚠️ Challenges to Keep in Mind
EWC is powerful, but not without limitations:
- Requires known task boundaries (e.g., when one interest ends and another begins)
- Doesn’t scale well to dozens or hundreds of tasks unless optimised
- Needs efficient Fisher Matrix approximation for large models
That said, in most web personalisation setups — where tasks arrive in a predictable stream (e.g., based on time or session) — EWC works incredibly well.
🚀 Final Thoughts
Web personalisation isn’t just about responding to the latest click. It’s about understanding users over time — respecting their shifting interests while retaining the essence of what makes their preferences unique.
Elastic Weight Consolidation is a vital part of that puzzle.
By protecting what’s important while allowing flexibility, EWC helps build recommendation engines that are not only adaptive, but also resilient, personal, and deeply user-centric.