TSEditor: Interactive Time Series Editing for Privacy Preservation

Paper Title

TSEditor: Interactive Time Series Editing for Privacy Preservation

Publication Info

• Topic area: Privacy-preserving techniques for time series data
• Keywords: Time series, privacy preservation, interactive editing, visual analytics, data utility, temporal patterns, anomaly detection, user study, multivariate data, data sharing

Background and Problem

• Problem / challenge: Time series datasets pose significant privacy risks due to identifiable temporal patterns, which can lead to re-identification or sensitive information leakage. Existing privacy-preserving methods are either too rigid, introduce excessive noise, or lack user control and interpretability.
• Significance: Addressing privacy risks in time series data is critical for enabling secure data sharing while retaining analytical utility, especially in domains like healthcare, energy, and transportation.
• Motivation and related work: Prior methods include value perturbation, temporal perturbation, and synthesis-based approaches, but they lack fine-grained, interactive control for users. Visual analytics systems for time series data exist but are not designed for privacy risk detection and mitigation. This paper addresses the gap by introducing an interactive, user-driven approach.

Solution

• Proposed approach: TSEditor, an interactive system for identifying and mitigating privacy risks in time series datasets through visual analytics and direct manipulation tools.
• Novelty:
  1. A taxonomy of temporal privacy risks (time, magnitude, and pattern) and corresponding mitigation strategies.
  2. A visual analytics system with six coordinated views for risk identification, editing, and evaluation.
  3. Development of six interactive editing operations (time perturbation, magnitude perturbation, and pattern substitution).
  4. Multi-faceted evaluation demonstrating effectiveness, usability, and data utility preservation.
• Procedure and key techniques:
  • Identification: Use coordinated views (ranking, stream, radial, and pattern) to detect privacy risks.
  • Editing: Apply interactive operations (e.g., Move-x, Move-y, Curve, Clone, Removal) to mitigate risks.
  • Evaluation: Assess the impact of edits on privacy risks and data utility using visual feedback and downstream task performance.

Results

• Concrete findings:
  • Minimal impact on data utility: 1.12% MAE and 0.89% MSE changes in prediction tasks; no significant classification performance drop.
  • High usability: SUS score of 85.5 and low NASA-TLX workload (M = 17.97).
  • Efficient editing: Average of 2.4 operations per task with broad adoption of editing tools.
• Advantage over baselines: Outperformed automated differential privacy methods by preserving data utility while providing user control and interpretability.
• Experiments / evaluation:
  • Two case studies: REFIT electrical load and Capture-24 activity tracker datasets.
  • Expert interviews with four domain experts.
  • User study with 12 participants, demonstrating task completion efficiency and usability.
  • Model evaluation to verify data utility preservation.
• Limitations and future work:
  • Limited support for multivariate relationships and cross-variable dependencies.
  • Workflow interruptions due to separate editing and visualization environments.
  • Need for real-time utility feedback and privacy-aware benchmarks.

Summary

TSEditor is an interactive system designed to address privacy risks in time series datasets while preserving analytical utility. It introduces a taxonomy of temporal privacy risks and provides six coordinated views and editing operations for targeted risk mitigation. Evaluations, including case studies, expert interviews, and a user study, demonstrate its effectiveness, usability, and minimal impact on data utility. Future work will focus on multivariate dependencies, seamless workflows, and real-time feedback to further enhance the system's capabilities.