TSQAgent: Rating Time Series Data Quality via Dedicated Agentic Reasoning
This paper introduces TSQAgent, a novel agentic reasoning framework that enhances large language models' ability to assess time series data quality by dynamically identifying relevant dimensions and performing grounded quantitative comparisons, thereby improving downstream data selection and model performance.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a chef trying to make a delicious soup. You have two buckets of vegetables in front of you. One bucket has fresh, crisp carrots and perfect potatoes. The other bucket has some rotten bits, muddy water, and vegetables that are cut into weird, inconsistent shapes.
If you just glance at the buckets, you might think they look "okay." But to make a great soup, you need to know exactly which bucket is better and why. Is the first bucket better because the carrots are fresher? Or is the second bucket actually better because the potatoes are larger, even if the carrots are a bit bruised?
This is exactly the problem the paper TSQAgent is trying to solve, but instead of soup, it's dealing with Time Series Data (like stock prices, weather reports, or heart rate monitors).
The Problem: AI is Bad at "Sniffing" Data
The researchers found that current Artificial Intelligence (specifically Large Language Models, or LLMs) is surprisingly bad at judging the quality of this data.
- The "Guessing Game": When asked to compare two sets of data, the AI often guesses the wrong reasons. It might say, "This data is bad because of the color," when the real issue is that the numbers are missing or the pattern is broken.
- The "Math Struggle": Even if the AI knows what to look for, it's terrible at doing the actual math to prove it. It tends to "hallucinate" (make things up) rather than calculating the exact difference in noise or trends.
To prove this, the authors built a test called TSQBench. It's like a standardized test for AI, where they take two sets of data, secretly ruin one in specific ways (like adding static noise or removing chunks of data), and ask the AI to spot the difference. The results showed that even the smartest AI models were struggling, often getting the "why" wrong and the "which is better" answer only slightly better than a coin flip.
The Solution: The "Three-Headed Detective" (TSQAgent)
To fix this, the authors created a new system called TSQAgent. Instead of asking one AI to do everything, they built a team of three specialized "agents" (AI roles) that work together like a detective squad:
The Perceiver (The Detective with a Magnifying Glass):
- Job: This agent looks at the two data sets and decides what to investigate.
- The Trick: Instead of checking every single thing (which wastes time and leads to confusion), this agent is trained to ignore the noise and focus only on the most important clues, like "Is there missing data?" or "Is the pattern consistent?" It learns to pick the right "quality dimensions" to check.
The Inspector (The Lab Technician):
- Job: Once the Perceiver says, "Check the noise level," the Inspector takes over.
- The Trick: This agent doesn't just guess. It is equipped with external tools (like a calculator or a microscope). It runs actual mathematical tests on the data to get hard numbers. If the Perceiver says "Check the trend," the Inspector uses a tool to measure the slope and says, "Okay, Series A goes up 5%, Series B goes up 2%." This stops the AI from making things up.
The Adjudicator (The Judge):
- Job: This agent collects the reports from the Inspector.
- The Trick: It acts like a judge in a courtroom. It looks at all the evidence. If the evidence is weak or contradictory, it sends the case back to the Inspector to "re-check" or asks the Perceiver to look for new clues. Once everything is clear, it makes the final verdict: "Series A is higher quality," along with a confidence score and a clear explanation.
The Results: Smarter Data, Better Soup
The researchers tested this new "Three-Headed Detective" system on real-world data (like electricity usage, traffic patterns, and medical records).
- Better Judgments: The system became much better at spotting exactly why data was bad and which data was better.
- Efficiency: Because the system was so good at picking the right data, they could throw away 25% of the "bad" data and still train their AI models just as well as if they had used 100% of the data.
- The "Magic" Stat: In one experiment, they used their system to select only 75% of the available data to train a massive AI model. The result? The model performed just as well as if it had been trained on 100% of the data.
Summary
In short, the paper says: "Current AI is bad at judging data quality because it guesses instead of calculating. We built a team of AI agents that work together—one to pick the right clues, one to do the math with tools, and one to make the final call. This system helps us pick the 'fresh vegetables' from the 'rotten ones,' allowing us to build better AI models using less data."
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.