We demonstrate how to use interactive visualisations to explore model residuals and structure in both the high-dimensional space and its 2-D embedding. These interactive diagnostics allow users to link views across dimensions, enhancing interpretability and supporting more informed decision-making during model evaluation.
Step 1: Augmenting data with model errors
We begin by computing model errors using the augment()
function. This step appends prediction errors from both the
high-dimensional and 2-D models to the original dataset.
model_error <- augment(
highd_data = scurve,
model_highd = scurve_model_obj$model_highd,
model_2d = scurve_model_obj$model_2d
)Step 2: Combining all components
Next, we prepare a comprehensive data frame that brings together the high-dimensional data, 2-D embeddings, model predictions, and error metrics. This combined object serves as the foundation for linked interactive plots.
df_exe <- comb_all_data_model_error(
highd_data = scurve,
nldr_data = scurve_umap,
model_highd = scurve_model_obj$model_highd,
model_2d = scurve_model_obj$model_2d,
error_data = model_error
)We also extract edge data (e.g., triangulations) used to define neighborhood relationships in the 2-D space.
edge_data <- scurve_model_obj$trimesh_dataStep 3: Generating linked interactive plots
Finally, we use the show_error_link_plots() function to
produce interactive linked views. These plots allow users to examine
prediction errors spatially, investigate neighborhood relationships via
edge links, and interactively select and highlight points across
plots.
show_error_link_plots(point_data = df_exe, edge_data = edge_data)This interactive visualisation bridges the gap between model diagnostics and exploratory data analysis, enabling a richer understanding of model performance and structural preservation in dimension reduction.