visualisation

In 1957, linguist J. R. Firth observed that 'you shall know a word by the company it keeps'. That principle — words that co-occur share meaning — is the foundation on which all of generative AI was built, from early Latent Semantic Analysis to today's trillion-parameter Transformers. This post traces the lineage with three interactive LSA-to-PCA visualisations in R (Reuters newswire, State of the Union addresses and IMDB reviews), showing where simple co-occurrence models succeed, where they fail and why scale alone turned a modest insight into the technology behind ChatGPT. It then examines why LLMs are optimised for fluency rather than truth — hallucinations are a structural consequence, not a bug to be patched — and argues that careful prompt engineering is the best tool we have for steering a fundamentally heuristic machine.

A custom R function to create ggplot2 visualizations of power curves generated by the simr package's powerCurve function for mixed-effects models.

Custom functions that extend sjPlot to discretise continuous variables in interaction plots into equally-sized intervals (deciles or sextiles) that include minimum and maximum values, with legend counts showing sample sizes per level.

Whereas the direction of main effects can be interpreted from the sign of the estimate, the interpretation of interaction effects often requires plots. This task is facilitated by the R package sjPlot. For instance, using the plot_model function, I plotted the interaction between a continuous variable and a categorical variable. The categorical variable was passed to the fill argument of plot_model. library(lme4) #> Loading required package: Matrix library(sjPlot) #> Install package "strengejacke" from GitHub (`devtools::install_github("strengejacke/strengejacke")`) to load all sj-packages at once!

A custom R function to create ggplot2 visualizations of fixed effects from models refitted with multiple optimizers using lme4's allFit function, enabling visual assessment of convergence validity in mixed-effects models.

When a model has struggled to find enough information in the data to account for every predictor---especially for every random effect---, convergence warnings appear (Brauer & Curtin, 2018; Singmann & Kellen, 2019). In this article, I review the issue of convergence before presenting a new plotting function in R that facilitates the visualisation of the fixed effects fitted by different optimization algorithms (also dubbed optimizers).

Whereas the direction of main effects can be interpreted from the sign of the estimate, the interpretation of interaction effects often requires plots. This task is facilitated by the R package sjPlot (Lüdecke, 2022). In Bernabeu (2022), the sjPlot function called plot_model served as the basis for the creation of some custom functions. One of these functions is alias_interaction_plot, which allows the plotting of interactions between a continuous variable and a categorical variable.

Whereas the direction of main effects can be interpreted from the sign of the estimate, the interpretation of interaction effects often requires plots. This task is facilitated by the R package sjPlot (Lüdecke, 2022). In Bernabeu (2022), the sjPlot function called plot_model served as the basis for the creation of some custom functions. Two of these functions are deciles_interaction_plot and sextiles_interaction_plot. These functions allow the plotting of interactions between two continuous variables.

This post presents a run-through of a Bayesian workflow in R. The content is *closely* based on Bernabeu (2022), which was in turn based on lots of other references, also cited here.