💻 computer science

TimeLAVA: Learning-Agnostic Data Valuation for Time Series

TimeLAVA is a learning-agnostic framework for time series data valuation that utilizes a novel Selective Wavelet-based Wasserstein discrepancy to efficiently compute robust, model-independent sample scores that capture temporal dependencies and distributional shifts without requiring model training.

Original authors: Wenqin Liu, Weizhi Quan, Aoqi Zuo, Erdun Gao, Vu Nguyen, Dino Sejdinovic, Howard Bondell, Mingming Gong

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wenqin Liu, Weizhi Quan, Aoqi Zuo, Erdun Gao, Vu Nguyen, Dino Sejdinovic, Howard Bondell, Mingming Gong

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 the perfect soup. You have a massive pot of ingredients (your time series data), but some are fresh and delicious, some are spoiled, and some are just weirdly out of place. If you throw everything in blindly, your soup will taste bad. You need a way to taste-test every single ingredient to decide which ones to keep and which to toss.

This is exactly what the paper TimeLAVA does, but instead of soup, it's dealing with time series data—streams of information that change over time, like heart rate monitors, stock prices, or factory sensors.

Here is a simple breakdown of how TimeLAVA works and why it's special:

1. The Problem: Why Old Methods Fail

Most existing ways to judge data quality are like trying to judge a movie by looking at individual frames without seeing the story.

The "Model-Dependent" Trap: Some methods require you to build a specific AI model first to see which data helped it learn. This is like hiring a critic to taste your soup after you've already cooked it. It's slow, expensive, and the critic might only like spicy food (biased to one model).
The "Static" Trap: Other methods assume data points are independent, like individual apples in a basket. But time series data is more like a river. The water at one moment depends on what happened a second ago. If you treat a river like a pile of rocks, you miss the current, the flow, and the patterns.

2. The Solution: TimeLAVA (The "Smart Taster")

TimeLAVA is a new tool that values data without needing to train a model first (it's "learning-agnostic"). It acts like a super-smart taster that compares your "Evaluated" data against a "Reference" data (a known good standard).

It uses two main tricks to do this:

Trick A: The "Wavelet" Flashlight

Imagine looking at a song. A standard tool (Fourier Transform) tells you what notes are in the song, but not when they happen. It's like knowing a song has a drum beat but not knowing if the drummer started at the beginning or the end.

TimeLAVA's Wavelets: These are like a flashlight with a zoom lens. They can look at the whole song (long-term trends) and then zoom in to see a specific drum hit (a sudden spike or glitch) at a precise moment in time. This allows TimeLAVA to spot both long-term patterns and sudden, short-lived anomalies.

Trick B: The "Selective Match" (Unbalanced Transport)

Imagine you are trying to pair up socks from two different piles.

Old Methods: They force you to pair every sock, even if one is a bright neon green and the other is a dull grey. They force a match, which creates a "bad pair" and ruins your score.
TimeLAVA's Unbalanced Transport: This method is smarter. It says, "If these two socks are too different, don't force a match." It allows the weird, mismatched socks to stay unmatched. This prevents one weird outlier (a single bad sock) from ruining the score of the whole pile. It's robust against "regime shifts" (when the whole style of data changes) and random noise.

3. How It Gives a Score

Once TimeLAVA compares your data to the reference using these tricks, it calculates a value score for every tiny segment of time.

High Score: "This segment fits perfectly with the reference. It's high-quality data."
Low (Negative) Score: "This segment is weird. It doesn't match the reference. It might be an anomaly, noise, or a corrupted label."

Crucially, it does this without training a single AI model. It just looks at the math of how well the data matches up.

4. What It Can Do (Based on the Paper's Experiments)

The authors tested TimeLAVA on real-world data and showed it works better than existing methods in three specific areas:

Spotting Anomalies (The "Smoke Alarm"):
- Scenario: A heart monitor shows a sudden, impossible spike.
- Result: TimeLAVA correctly identifies the exact moment the spike happened as "low value" (bad data), while ignoring the normal, healthy beats. It found these errors better than other methods.
Cleaning the Data (The "Data Pruner"):
- Scenario: You have a huge dataset for training a model, but 20% of it is corrupted with noise (like static on a radio).
- Result: TimeLAVA can rank the data segments. If you throw away the "lowest value" segments, the model trained on the remaining "high value" data performs much better. It successfully identified and removed the "rotten apples."
Finding Bad Labels (The "Proofreader"):
- Scenario: You have medical data where doctors labeled patient conditions, but some labels are wrong (e.g., a "healthy" label on a sick patient).
- Result: TimeLAVA can spot these mistakes. It noticed when the label didn't match the pattern of the data, especially when the errors happened in specific patterns (like a sensor failing every 10 minutes).

Summary

TimeLAVA is a new, model-free way to grade time series data. Instead of forcing a square peg into a round hole, it uses wavelets to see details at different speeds and selective matching to ignore impossible pairings. This lets it accurately tell you which parts of your data are gold and which parts are garbage, helping you build better, more reliable AI systems without the heavy cost of training models just to check the data quality.

Technical Summary: TimeLAVA

Problem Statement

Time series data underpins critical decision-making systems in healthcare, finance, and industrial monitoring. However, identifying the intrinsic value of specific temporal segments—distinguishing informative patterns from anomalies, noise, or distributional shifts—remains a fundamental challenge. Existing data valuation methods suffer from two primary limitations:

Model-Dependency: Approaches like Influence Functions or TimeInf are tied to specific model architectures and training procedures, limiting their generalizability.
Inapplicability to Sequential Data: Learning-agnostic methods designed for independent and identically distributed (i.i.d.) data (e.g., LAVA, SAVA) fail to capture temporal dependencies, multi-scale patterns, and non-stationary dynamics inherent to time series.

The core challenge is to develop a valuation framework that is learning-agnostic (independent of specific models), robust to non-stationarity and outliers, and capable of capturing multi-scale temporal structures without requiring expensive model retraining.

Methodology: The TimeLAVA Framework

The authors propose TimeLAVA, a framework that values temporal segments based on their marginal contribution to minimizing the distributional discrepancy between evaluated data and a reference distribution. The methodology consists of three key components:

1. Multi-Scale Temporal Representation (Wavelets)

To address the limitations of Fourier transforms (which lack temporal locality) and standard metrics, TimeLAVA employs the Discrete Wavelet Transform (DWT).

Mechanism: Segments are transformed into multi-scale wavelet coefficients, preserving both time and frequency characteristics.
Distance Metric: A Wavelet Distance ( $d_{wav}$ ) is defined as the $L_1$ norm of the difference between wavelet coefficients. This allows the framework to capture transient events, localized dynamics, and sustained trends simultaneously.

2. Selective Wavelet-based Wasserstein (WSW) Discrepancy

To handle non-stationarity and outliers, the framework integrates wavelet distances with Unbalanced Optimal Transport (UOT).

Rationale: Standard Optimal Transport (OT) enforces strict mass conservation, forcing even dissimilar segments (e.g., regime shifts or outliers) to match, which inflates discrepancy estimates.
Solution: TimeLAVA uses UOT, which relaxes mass conservation via KL-divergence penalties. This enables selective matching, allowing dissimilar segments to remain partially unmatched rather than forced into high-cost alignments.
Cost Function: The cost matrix combines the wavelet distance between segments and a term for label consistency (if labels are available), controlled by a weight parameter $c$ .

3. Efficient Valuation via Dual Sensitivity Analysis

Instead of retraining models, TimeLAVA computes segment values efficiently through the dual formulation of the UOT problem.

Valuation Logic: The value of a segment is derived from its dual potential in the optimal transport plan. Segments with low dual potentials (easy to match) receive positive values, while those with high potentials (hard to match) receive negative values.
Point-wise Scores: For applications requiring fine-grained resolution (e.g., pinpointing anomaly timestamps), segment-level values are aggregated using a sliding window approach to produce point-wise scores.
Complexity: The approach avoids model training, with computational complexity dominated by wavelet transforms and entropy-regularized UOT solving ( $O(nmL + nm\log(1/\epsilon))$ ).

Key Contributions

TimeLAVA Framework: A learning-agnostic valuation framework specifically designed for time series, capable of capturing complex, evolving dynamics without model training.
Selective Wavelet-based Wasserstein (WSW) Discrepancy: A novel metric combining multi-scale wavelet transforms for temporal localization and unbalanced optimal transport for robustness to distributional shifts and outliers.
Theoretical Guarantees:
- Bounded Sensitivity: Proven that WSW exhibits bounded sensitivity to outlier contamination, distinguishing isolated anomalies from systemic regime shifts (Theorem 5.1).
- Generalization Bound: Established a theoretical link between the WSW discrepancy and model-agnostic generalization, proving that minimizing this discrepancy improves downstream task performance for any well-behaved model (Theorem 5.3).
Empirical Validation: Demonstrated superior performance across diverse real-world datasets in three distinct tasks:
- Anomaly Detection: Outperformed state-of-the-art methods (e.g., TimeInf, Isolation Forest, Deep Learning detectors) in identifying anomalous segments.
- Data Pruning & Selection: Successfully identified and removed low-quality/corrupted segments, maintaining or improving forecasting performance (measured by $R^2$ ) while random or other baselines degraded.
- Temporal Label Noise Detection: Effectively identified time-varying label corruption patterns (periodic, decay, growth) in healthcare datasets, outperforming baselines in F1 scores.

Significance and Claims

The paper claims that TimeLAVA provides a principled, learning-agnostic approach to quantifying intrinsic data quality in time series. By decoupling valuation from specific model training, it offers a scalable and generalizable solution for data curation.

Practical Impact: The method enables robust anomaly identification, efficient data pruning to reduce training costs, and detection of noisy labels in dynamic environments.
Theoretical Grounding: Unlike heuristic methods, TimeLAVA is grounded in optimal transport theory, with formal proofs linking distributional discrepancy to generalization bounds and robustness.
Limitations Acknowledged: The authors note that the framework's effectiveness relies on the representativeness of the reference set; biased references may undervalue rare patterns. Additionally, the current implementation processes channels independently, not explicitly modeling cross-channel dependencies. Future work is suggested to address scalability via mini-batch strategies and to extend the framework to active learning and classification tasks.

The code is publicly available, facilitating reproducibility and further research in time series data curation.