Model literacy: the hidden factor boosting visual analytics accuracy

Visual analytics (VA) tools help people explore data through interactive visualizations, but evaluating their effectiveness is tricky. Existing evaluation methods focus on visualization literacy—how well users read charts. Now researchers from a team including Lei Xia argue we need a second factor: "model literacy," or how well users understand the analysis model driving the visualization. Their paper, posted on arXiv (2608.18721), tests this concept with two common dimensionality-reduction models: PCA and t-SNE.

In a controlled study where participants performed tasks on multidimensional data using VA tools built on either model, the team found a positive correlation between model-task accuracy (measuring knowledge of the model's behavior) and VA-task accuracy. That correlation was stronger for PCA than for t-SNE, likely because PCA's linear projections are less directly readable in the visualization—so users needed more model understanding to compensate. Interestingly, completion time showed no stable efficiency gain: sometimes model knowledge sped up tasks, sometimes it added interpretive overhead, suggesting the intuitiveness of the model plays a critical role. The authors argue this opens a new direction for VA evaluation and calls for more rigorous model-literacy assessment instruments.