Comparison Table: 2D Projection vs Unity 3D

Overview

This page compares three conditions for visualizing the same 64-word GloVe embedding dataset: a noisy 2D PCA projection (using low-variance dimensions 3 and 5), a clean 2D PCA projection (using highest-variance dimensions 1 and 2), and an interactive Unity 3D environment using UMAP layout on the Meta Quest 3 headset. Data is drawn from a user study conducted March 10, 2026 with 9 participants.

Feature Comparison

Quantitative Results (n=9)

Cluster Definition Rating (1-5)

Visualization Preference (most chosen per task)

Key Finding: Noisy Projection Inverts Similarity Judgment

The most striking quantitative result from the study is the doctor/nurse/engineer similarity judgment. In the underlying GloVe embedding space, 'doctor' is substantially closer to 'nurse' than to 'engineer'. However, in the noisy 2D projection (PCA dims 3 & 5), the spatial layout places doctor visually closer to engineer, misleading 78% of participants into the wrong answer. In the clean 2D and Unity 3D conditions, the correct answer dominated at 89% and 78% respectively.

This demonstrates that visualization quality is not a cosmetic concern — a poorly chosen 2D projection can actively mislead users about the structure of the underlying data.

Qualitative Themes

What helped most in Unity 3D

• Physical navigation allowed participants to resolve ambiguous spatial relationships that were unclear in 2D.

• Depth perception made cluster separation more intuitive — participants could tell at a glance which words were truly close vs merely overlapping in projection.

• The ability to approach specific words from different directions helped users build a mental model of the embedding space.

What worked against Unity 3D

• Labels only visible on hover required users to point at each sphere individually, which was slow for systematic exploration.

• The info panel sometimes appeared behind the user or at an awkward world-space angle.

• Two participants noted that the 3D space showed more inter-cluster overlap than expected, which was initially surprising (though arguably more honest than the clean 2D projection).

Recommendations

Best for: Use Unity 3D when:

• The primary goal is building intuition about cluster structure and spatial relationships

• Users need to compare distances between points across multiple dimensions

• Time and hardware are not constraints

Best for: Use clean 2D when:

• Quick overview of gross cluster structure is the goal

• Hardware access is limited or not possible

• The visualization will be shared in a paper, poster, or printed format

Avoid if: Avoid noisy 2D when:

• Accurate similarity or distance judgment is required from the visualization

• Users are not experts who can mentally adjust for projection distortion