Janith Wanniarachchi

Research

I am a PhD candidate in the Department of Econometrics & Business Statistics at Monash University, supervised by Dianne Cook, Patricia Menendez, Thiyanga Talagala and Kate Saunders. My research asks how we can see what explainable AI methods are telling us, and whether simpler, interpretable models can stand in for complex ones.

PhD thesis

Explaining AI through visual explanations and efficient interpretable models

Local explainers such as LIME, SHAP, Anchors and counterfactuals each describe a different facet of a model, and they often disagree. The thesis places their explanations back in the data space, where they can be explored with high-dimensional visualisation such as tours. It rebuilds two explainers in R so that every intermediate step can be inspected, and it shows that small, well-chosen models can rival larger ones.

Read the thesis Source code

Projects

Visual explanations of the explainers

Draws SHAP, LIME, counterfactuals and Anchors geometrically in the data space so they can be compared side by side. In the same model they often name different features as most important.

With D. Cook, K. Saunders, P. Menendez & T. Talagala

Chapter · Blog post · useR! 2026 talk

Random seeds, stability & parsimony

Single-layer neural networks show that simple models can match complex ones but are harder to find: their accuracy swings with the random seed. The project comes with a Shiny teaching tool.

Neural networks · Shiny · teaching

Chapter · Blog post

kultarr: Anchors, revisited

A pure-R implementation of Anchors that you can debug and visualise. It uses S7 classes, memoised rewards and parallel search, and includes helpers for drawing anchors in high dimensions.

R package · in development

Chapter · Docs · Blog post

kumquat: a simpler LIME

A transparent take on LIME for tabular data. It returns the perturbations, predictions and local GLM, so you can audit exactly how an explanation was built.

R package · CRAN

Chapter · Docs · CRAN

sillysplines: drawn decision boundaries

An interactive widget for sketching a linear-spline decision boundary and downloading the coordinates as a synthetic classification dataset, for teaching and model testing.

R package · htmlwidget

Chapter · Docs

scatteR: inverse scagnostics

My undergraduate thesis. It generates scatterplots that match target scagnostic measures by placing points with simulated annealing, which gives data for testing visual and ML methods.

R package · CRAN · useR! 2022

Preprint · CRAN · GitHub

Publications & presentations

  • Wanniarachchi, J. (2026). Visual Explanations of XAI Explainers. useR! 2026. Slides & code
  • Wanniarachchi, J. (2023). Seeing the Smoke Before the Fire: Visualising Spatiotemporal Indicators of Bushfires through Counterfactual Explanations of Machine Learning Models. 12th Conference of the Asian Regional Section of the IASC (IASC-ARS), Macquarie University, Sydney. Outstanding Young Researcher / Student Contributed Oral Presentation Award. Slides
  • Wanniarachchi, J. C., & Talagala, T. S. (2022). scatteR: Generating instance space based on scagnostics. arXiv:2209.06682 [stat.CO]. Presented at useR! 2022. doi:10.48550/arXiv.2209.06682 · Talk
  • Karunanayake, V. J., Wanniarachchi, J. C., Karunanayake, P. N. K., & Rajapaksha, U. U. S. (2022). Intelligent System to Verify the Effectiveness of Proposed Teacher Transfers Incorporating Human Factors. 2nd International Conference on Advanced Research in Computing (ICARC), IEEE. doi:10.1109/ICARC54489.2022.9754089

Recognition