Ordered Treemap Layouts ======================= Slice-and-dice creates long, thin rectangles. Clustered algorithms and squarified algorithms try to make aspect ratio near 1. But: layout changes abruptly when data changes (hinders reading, selection, comparison, and recall). Ordering is no longer preserved. Just eliminate the sorting step. Ordered Algorithm ----------------- Choose a pivot element, P. Place rectangle R_P along top or left side, whichever is longest. Place items ordered before P in rectangle R_1 (above P). Place items after P in R_2 (to the right of P), R_3 (below P) to optimize aspect of R_P. Recurse on R_1, R_2, R_3. .----------------. | | | | | | | R_1 | | | | | | | | | |----------------| | | | | R_P | R_2 | | | | |----------------| | | | | | R_3 | | | | | .----------------. Strip Treemap Algorithm ----------------------- Arrange items in horizontal strips. Goal is low average aspect ratio, not uniform strip height. How to choose pivot? -------------------- Choose largest item. Choose middle-sized item. Choose so that groups of items before and after pivot are roughly the same size. Choices: slice-and-dice; pivot by middle; pivot by split size; strip treemap; cluster; squarified. Evaluation ---------- Consider: *aspect ratio* (average for all leaf rectangles, using max(w, h)/min(w, h)); *change* (rectangle = (x, y, w, h), use Euclidean distance in R^4); *disorder* (draw a sequence of line segments connecting rectangles in order and measure percentage of instances where line segments change direction). Slice-and-dice has bad aspect, good ordering. Squarified and clustered have good aspect, bad ordering. Strip treemaps are a good compromise. Measures of disorder correlate well to reaction times for selection tested on people. Mixed Layouts ------------- Use different layouts at different levels. At top level, use most stable; at leaf nodes, optimize aspect ratio.