Projection inspector: Assessment and synthesis of multidimensional projections

Multidimensional Data Exploration by Explicitly Controlled Animation

“…Within the scope of multidimensional visualization, parameter-space animations include rolling-the-dice [39], where the user controls the plane on which the multidimensional data is projected; the grand tour [40], where a large sequence of 2D projections are displayed from a multidimensional dataset in a flip-book manner (the parameters P controlling projection-plane orientation in nD being varied randomly); the class tour [41], which refines the grand tour so as to generate projections which preserve class separation of the data points; combinations of the grand tour with projection pursuit [42] (the parameters P controlling the projection-plane orientation being varied along the derivatives of the so-called projection pursuit index, so as to drive the tour through interesting projections); and drawing faded trails that connect two consecutive views in a tour to give a feeling of how the projection plane changed in between [43]. A limitation of most grand tour techniques is that they only handle attribute-based projections [44].…”

“…One can show a history of views V(P i ) obtained for various parameter settings P i used in the exploration so far. The history can be shown as a linear, grid-like, or hierarchical set of thumbnails depicting V(P i ), a metaphor also called 'projection board' [44]. By clicking on the desired thumbnail, the user can go back to the corresponding state P i and associated view V(P i ), and continue exploration from there.…”

“…Most projection pursuit variants also fall into this category [42,46], every view V i being one deemed 'interesting' from the perspective of the projection pursuit index, a metric which typically measures the distance from a given projection to an uninteresting generic Gaussian-like scatterplot. However, a limitation of most projection pursuit variants is that, just like grand tour variants, they require attribute-based projections [44]. Projection pursuit methods have been recently enhanced by limiting the number of interesting views to n/2 [47].…”

“…While powerful as an interaction mechanism, TPP limits itself to only linear projections. Recently, ProjInspector use a preset controller, where the k presets correspond to user-chosen DR projections V i (D), 1 ≤ i ≤ k. When one drags the point of interest in the controller, the tool generates a view V(D) that blends all V k (D) by means of mean values coordinates interpolation [44]. ProjInspector is arguably the closest technique to the one we present here, as such, we will discuss the similarities and differences in detail in Section 5.…”

Visualização e exploração de dados multidimensionais na web

“…Para validar a qualidade da projeção e o visualizar o impacto de erros causados pelo processo de perda de dimensionalidade, foi implementada a métrica de preservação de vizinhança suave ou Smooth Neighborhood Preservation (SNP) (PAGLIOSA et al, 2015), descrita a seguir. falso positivos).…”