Scaling Time Series Classification via XAI-Driven Data Reduction

Imagine you are a detective trying to solve a mystery, but instead of a few clues, you are handed a mountain of evidence: thousands of pages of notes, hundreds of hours of surveillance footage, and a room full of people talking at once. This is the daily reality for computers trying to learn from Time Series Data. Think of this data as a long, continuous story told by numbers—like a heartbeat monitor, a stock market ticker, or a weather sensor recording every second for days. To make sense of these stories, we use Machine Learning, a type of computer program that learns patterns to make predictions, like guessing if a heart rhythm is healthy or if a storm is coming.

However, there is a catch. The more detailed the story (the longer the time series) and the more characters involved (the more sensors or "channels" recording data), the harder it is for the computer to read it. It's like trying to read a library of books all at once; the computer gets overwhelmed, runs out of memory, and crashes. This is where Explainable AI (XAI) comes in. Usually, XAI is like a translator that helps us understand why a computer made a decision. But what if we could flip the script? What if we used that translator not just to explain the answer, but to tell us which parts of the story actually matter, so we can throw away the rest before the computer even tries to read it? That is the big question this paper explores: Can we use the "why" to make the "how" much faster and lighter?

The Detective's Shortcut: drXAI

In this paper, the researchers introduce a clever new tool called drXAI (Data Reduction with XAI). Think of it as a super-smart editor for time series data. Its job is to look at a massive, messy dataset and say, "Hey, 90% of this is just noise! Let's cut it out so the computer can focus on the good stuff."

Usually, when we want to teach a computer to classify time series (like telling the difference between a running dog and a sleeping dog based on sensor data), we use powerful, complex models. These models are like brilliant but gluttonous chefs: they need huge amounts of ingredients (data) and a massive kitchen (computer memory) to cook a meal. If the ingredients are too many, the kitchen explodes (the computer runs out of memory).

The authors realized that while we have these brilliant chefs, we haven't been very good at filtering the ingredients before they get to the kitchen. So, they built a two-step process using drXAI:

1. The Quick Taste-Test: First, they use a fast, lightweight model called Hydra (think of it as a quick-tasting sous-chef) to look at a small sample of the data. This sous-chef is fast enough to run on a graphics card without breaking a sweat.
2. The "Why" Analysis: Next, they use Explainable AI techniques. Instead of just asking the sous-chef "What is this?", they ask, "Why did you think this was a dog?" The AI generates a map of "attribution," highlighting exactly which parts of the data (which time points or which sensors) were important for the decision.
3. The Great Cleanup: The researchers then take all these "importance maps" and combine them. They use a clever trick called an "elbow cut" (imagine bending a curve until it stops going up and starts flattening out) to automatically decide where to draw the line. They keep the top features that matter and toss the rest.

The Results: Cutting the Fat, Keeping the Muscle

The team tested this idea on both made-up data (where they knew exactly which parts were important) and real-world data (like whale sounds, military exercises, and heartbeats).

On the fake data: The results were like a magic trick. When they used traditional methods to pick out the important parts, they often grabbed the wrong clues or missed the right ones. But drXAI? It was spot on. In one test, it successfully picked out 19 out of 20 important channels, while other methods struggled to find even half. It knew exactly which sensors were talking and which were just making noise.

On real-world data: The impact was even more dramatic. The researchers found that drXAI could throw away 80% to 90% of the data (the time points or sensors) without hurting the computer's ability to make correct predictions. In fact, in many cases, the computer performed just as well on the tiny, cleaned-up dataset as it did on the massive, messy original one.

But the real "wow" moment came with the heavy-duty models. There is a very powerful model called ConvTran that is usually too hungry to run on huge datasets because it needs too much computer memory. The researchers tried to run it on a dataset with 44,000 time points, and it crashed immediately. However, after using drXAI to cut the data down, ConvTran was able to run successfully and make accurate predictions. It was like taking a giant, heavy elephant and teaching it to fit through a mouse hole by removing the extra weight it was carrying.

What This Means for the Future

The paper suggests that we don't have to choose between having a super-smart model and having a manageable amount of data. By using XAI not just to explain decisions, but to select the best data, we can make powerful AI scalable again.

The researchers also tested different "backgrounds" for their AI to compare against. They found that using a "zero" background (pretending missing data is just silence) wasn't as good as using a "prototype" background (pretending missing data looks like the average of that group). It's like trying to guess a missing word in a sentence: guessing it's a blank space is less helpful than guessing it's the most common word for that topic.

In short, drXAI shows that sometimes, to see the forest clearly, you don't need to count every single leaf. You just need to know which leaves are the most interesting, and let the rest fall away. This approach allows us to train smarter, faster, and on datasets that were previously too big to handle, all while keeping the computer from running out of memory.