TL;DR
This paper introduces RelaQ, an interactive system for querying multiple time series based on their relationships. It addresses the challenge of specifying and understanding complex, heterogeneous relationships between time series, offering a more intuitive and efficient approach than traditional scripting methods.
What Problem Does This Paper Address?
Analyzing multiple time series often involves understanding how different series relate to each other. For instance, in financial data, one might want to find periods where two stocks exhibit a local negative correlation despite an overall positive one. Current methods for querying time series primarily focus on individual series patterns or simple similarities, often requiring laborious scripting and lacking support for diverse, heterogeneous relationships. This gap makes it difficult for analysts to detect anomalies, validate hypotheses, and explore complex patterns across multiple time series datasets.
Methods
RelaQ tackles this problem through a novel interactive visual query system. Key aspects of its approach include:
- Formative Study: The researchers conducted a study with 11 domain experts to identify common time series relations (correlation, causality, similarity, lag, arithmetic, meta) and understand user pain points in querying.
- Interactive Query Interface: RelaQ provides an input panel where users can sketch trends and intuitively specify various relations between time series using a visual interface. This includes "timeboxes" for defining constraints on individual series (trends, values, names) and "relalinks" to connect them based on specified relationships.
- Heterogeneous Relation Matching: A core innovation is RelaQ's ability to handle multiple, diverse relation constraints simultaneously. It employs a graph-matching approach, modeling time series fragments as nodes and relations as links, to find matches within a dataset graph.
- Scalable Visualization: The system offers a multi-level visualization of query results, including a matrix view for comparing results and a time view for temporal exploration, aiding in the interpretation of complex query outcomes.
- Guidance and Exploration: RelaQ includes a guidance panel that suggests alternative timeboxes and relations, supporting users in exploring potential patterns beyond their initial queries.
Datasets
The paper does not explicitly mention specific datasets used for training RelaQ's models. However, it details the evaluation of RelaQ using:
- EEG Database: Used in Case Study 1 to analyze correlations between brain regions before and after alcohol consumption.
- PM2.5 Concentration Dataset: Used in Case Study 2 to analyze air pollution transmission patterns between cities in the North China Plain.
- Real Factory Dataset: A dataset with 142 time series and a length of 5,000 used in the user study.
Key Insights and Ideas
- The "Relation" is Key: The study highlights that understanding relationships between time series is as crucial as understanding individual series patterns for many real-world analysis tasks.
- Expert-Driven Taxonomy: Identifying and formalizing six distinct types of time series relations (correlation, causality, similarity, lag, arithmetic, meta) provides a structured vocabulary for these complex interactions.
- Bridging the Scripting Gap: RelaQ's visual and interactive approach significantly reduces the reliance on complex scripting, making advanced time series analysis more accessible.
- Visualization for Understanding: The multi-level visualization is critical for interpreting the often complex and numerous results of relation-driven queries, helping users discern meaningful patterns.
- Guidance for Exploration: Providing intelligent recommendations for further exploration is vital for users who may not have a precise target in mind, facilitating discovery.
Key Results
- Effectiveness and Usability: Case studies with domain experts (EEG analysis and air pollution analysis) demonstrated RelaQ's effectiveness in addressing real-world analytical questions.
- User Study Performance: A user study with 12 participants showed a high average task passing rate (0.98) and a good average System Usability Scale (SUS) score of 81.04, indicating strong usability and learnability.
- Efficiency: The average task completion time was 53 seconds, with an overall average of 7.8 minutes for all tasks, suggesting efficient interaction.
- Qualitative Feedback: Experts and users praised RelaQ's intuitive query formulation, flexible relation specification, and powerful visualization capabilities, while also suggesting areas for improvement.
Limitations and Future Work
- Scalability: While RelaQ performs well for medium-sized datasets, its time complexity (O(kN²M)) can lead to noticeable lags for very large datasets (e.g., 10⁸ time series).
- Relation Computation Flexibility: Currently, RelaQ uses standard computation methods for relations. Experts expressed a desire for optional or modifiable algorithms for more specialized analysis.
- Learnability vs. Advanced Functions: There's a trade-off between offering advanced functions (like flexible relation computation) and maintaining ease of learning.
- Integration with Intelligent Pattern Mining: The authors suggest integrating intelligent pattern mining models to enhance RelaQ's capabilities, particularly for optimizing query parameter selection and handling free-scale matching.
- Guidance Update: Some users found the guidance panel's lack of automatic updates inconvenient.
Summary updated