[Paper Review] Data Point Selection for Line Chart Visualization: Methodological Assessment and Evidence-Based Guidelines

Jonas Van Der Donckt, Jeroen Van Der Donckt+2 more|arXiv (Cornell University)|Apr 3, 2023

Data Visualization and Analytics4 citations

TL;DR

This paper proposes a metrics-based evaluation methodology for data point selection algorithms in line chart visualization, focusing on visual representativeness and a novel concept of visual stability. It evaluates EveryNth, MinMax, LTTB, and M4 across three toolkits, finding MinMax superior in visual stability and LTTB/MinMax best in representativeness, with empirical guidelines for data efficiency and line width trade-offs.

ABSTRACT

Time series visualization plays a crucial role in identifying patterns and extracting insights across various domains. However, as datasets continue to grow in size, visualizing them effectively becomes challenging. Downsampling, which involves data aggregation or selection, is a well-established approach to overcome this challenge. This work focuses on data selection algorithms, which accomplish downsampling by selecting values from the original time series. Despite their widespread adoption in visualization platforms and time series databases, there is limited literature on the evaluation of these techniques. To address this, we propose an extensive metrics-based evaluation methodology. Our methodology analyzes visual representativeness by assessing how well a downsampled time series line chart visually approximates the original data. Moreover, our methodology includes a novel concept called "visual stability", which captures visual changes when updating (streaming) or interacting with the visualization (panning and zooming). We evaluated four data point selection algorithms across three open-source visualization toolkits using our proposed methodology, considering various figure-drawing properties. Following the analysis of our findings, we formulated a set of evidence-based guidelines for line chart visualization at scale with downsampling. To promote reproducibility and enable the qualitative evaluation of new advancements in time series data point selection, we have made our methodology and results openly accessible. The proposed evaluation methodology, along with the obtained insights from this study, establishes a foundation for future research in this domain.

Motivation & Objective

• To address the lack of systematic evaluation of data point selection algorithms in time series visualization.
• To develop a metrics-based methodology that quantifies visual representativeness and introduces visual stability as a novel evaluation criterion.
• To assess the performance of four widely used algorithms—EveryNth, MinMax, LTTB, and M4—across different visualization toolkits and rendering configurations.
• To derive evidence-based guidelines for selecting optimal downsampling strategies in large-scale time series visualization.
• To promote reproducibility by open-sourcing the methodology, code, and experimental data.

Proposed method

• The methodology uses image-based metrics (e.g., SSIM, LPIPS, PSNR) to quantify visual representativeness, accounting for toolkit-specific rendering characteristics like line width and anti-aliasing.
• A novel OR-conv mask is introduced to enable targeted per-pixel aggregation, improving consistency across diverse time series templates.
• Visual stability is measured using two data-space metrics (e.g., L2 distance between downsampled sequences) to assess perceptual coherence during panning, zooming, or streaming.
• The evaluation is conducted across three open-source visualization toolkits (Plotly, D3.js, Matplotlib), with controlled variations in line width and anti-aliasing.
• The number of selected data points is used as a dependent variable to assess data efficiency, enabling fair comparison across algorithms.
• Empirical data is collected across 200k-point time series templates at three different sizes, with results aggregated across multiple zoom levels and configurations.

Experimental results

Research questions

• RQ1 How do different data point selection algorithms compare in visual representativeness across various visualization toolkits and rendering parameters?
• RQ2 To what extent does visual stability—defined as perceptual consistency during panning, zooming, or streaming—differ between algorithms?
• RQ3 What is the relationship between line width, canvas width, and the number of guaranteed alternations in determining the data efficiency elbow point?
• RQ4 Can the M4 algorithm achieve pixel-perfect representation in practice, especially under anti-aliasing?
• RQ5 Which algorithm offers the best trade-off between visual fidelity and perceptual stability for interactive time series exploration?

Key findings

• MinMax consistently outperformed other algorithms in visual stability, often by a significant margin, due to its localized selection strategy.
• LTTB and MinMax achieved the highest visual representativeness across most time series templates and rendering configurations, particularly under anti-aliasing and increased line width.
• The study empirically derived a formula relating line width, canvas width, and the number of guaranteed alternations to identify the performance elbow in data efficiency.
• M4 was found to be non-trivial to implement as pixel-perfect, and the study proves it is impossible to achieve pixel-perfect results under anti-aliasing due to rasterization effects.
• The proposed methodology revealed that visual stability is a critical but previously overlooked dimension in evaluating downsampling algorithms, especially for interactive visual analytics.
• The open-sourced codebase enables reproducibility and supports future user studies, with two toolkits (Plotly, D3.js) being web-integrable for survey-based evaluation.