useR! 2026

spinebil: Practical Diagnostics for Index Reliability in Exploratory Data Analysis

8 Jul 2026, 14:25

5m

Lightning Talk (5 minutes) Lightning Talks

Speaker

Tina Rashid Jafari (Monash University, Australia)

Description

• Author: Tina Rashid Jafari, Department of Econometrics and Business Statistics, Monash University, Australia, Email: tina.rashidjafari@monash.edu
• Title: spinebil: Practical Diagnostics for Index Reliability in Exploratory Data Analysis
• Primary Topic: Statistical Graphics/ Exploratory Data Analysis
• Keywords: Exploratory Data Analysis; Data Visualization; Projection Pursuit; Index Diagnostics; Statistical Graphics
• Abstract: Many R users rely on a single number index, such as an interestingness score, a dependence statistic, or a shape measure, to rank plots and guide exploration. Yet index behaviour is often taken for granted. In practice, indices can behave unexpectedly: their scales may be unclear, they may change with sample size, or they may react strongly to noise, making results hard to interpret and comparisons unreliable across datasets. Without diagnostics, it is difficult to know whether an index is measuring the intended feature or responding to artefacts.
  This talk presents spinebil, an R package designed to diagnose and improve the reliability of index functions used in exploratory analysis, screening, or optimisation. Any index can be evaluated under controlled conditions by simulating structured patterns and null data with no structure, repeating experiments across sample sizes, and generating diagnostic plots that summarise scale behaviour, variability, and sensitivity. The diagnostics focus on three practical questions: does the index stay low under null conditions and rise for the feature it claims to measure; how does it change with sample size; and how smoothly does the index degrade as noise is gradually added to signal. Percentile based null baselines can be estimated and used for calibration so values become more interpretable and comparable across datasets, and signal plus noise experiments can identify the noise level at which an index becomes indistinguishable from the null. A motivating example is financial time series: dependence or shape indices computed across rolling windows should be stable, comparable, and sensitive only when structure is distinguishable from noise. spinebil provides tools to test these properties before deployment.
• Presented at previous conferences?: No.
• Link to an external resource such as a GitHub repo or technical report:
• GitHub repo
• CRAN

If you used AI tools or services to support the preparation of this submission, please state the name and reason for using each of them.

ChatGPT was used only for grammar and wording edits to improve clarity.