TSQAgent: A New Framework for Time Series Data Quality Assessment Based on Agent Reasoning

This article introduces TSQAgent—an agent reasoning framework for time series data quality assessment. It addresses the shortcomings of existing Large Language Models (LLMs) in quality dimension identification and quantitative comparison through three collaborative roles: Perceiver, Inspector, and Arbiter. The framework has been validated on the TSQBench benchmark and 11 real-world datasets, improving assessment accuracy and translating into performance gains for downstream tasks.

Time series data is widely used in finance, IoT, meteorology, and other fields, but quality assessment is challenging due to the interweaving of multi-dimensional features (completeness, continuity, etc.). Traditional methods rely on manually predefined dimensions and rules/statistical indicators; existing LLM methods have two major issues: they depend on manual dimension definitions and cannot guarantee the identification of scenario-relevant dimensions; they only perform pure text reasoning and lack evidence-based quantitative comparison capabilities.

To evaluate LLM capabilities, the research team constructed the TSQBench benchmark, focusing on two core abilities:

1. Understanding and identifying relevant quality dimensions (e.g., continuity is needed for stock prediction, and outlier dimension for anomaly detection);
2. Quality comparison under specific dimensions. The results show that mainstream LLMs often miss key dimensions or introduce irrelevant ones in dimension identification, and their quality comparisons lack quantitative analysis, relying on surface feature judgments.

TSQAgent decomposes the assessment task into three roles:

1. Perceiver: Analyzes metadata, statistical features, etc., to generate a prioritized list of key quality dimensions, avoiding dimension explosion and omissions;
2. Inspector: Uses external tools to perform quantitative analysis on selected dimensions (e.g., missing rate for continuity, variance for smoothness) to provide a data foundation;
3. Arbiter: Weighted aggregation of results from each dimension, handles dimension trade-offs, generates comprehensive scores/conclusions, and has self-correction capabilities.

Experiments of TSQAgent on TSQBench and 11 real-world datasets yielded four key findings:

1. Significant improvement in dimension identification accuracy, especially with obvious advantages on complex high-dimensional data;
2. Substantial improvement in quantitative comparison capabilities—from qualitative description to quantitative analysis, leading to more consistent and interpretable conclusions;
3. Performance gains in downstream tasks: selecting data based on assessment results leads to better performance in prediction tasks;
4. Improved data efficiency: filtering low-quality data allows models to achieve better results with less data.

Technical significance: It proves that the agent reasoning framework can enhance the vertical domain capabilities of LLMs, and the three-role design provides a paradigm for other assessment tasks. Application prospects: Integration into data pipelines as a quality gate (scoring before data storage, generating reports); screening high-quality training data during the data selection phase to improve model efficiency.

Limitations: The external toolset covers statistical/time series analysis but needs to be extended to domain knowledge dimensions (e.g., financial compliance, medical clinical validity); it relies on LLM reasoning, which may fail in extremely complex problems. Future directions: Support for real-time stream data monitoring, adaptive dimension weight learning, and multi-modal time series applications.