Non-Negative Matrix Factorization: a Cone of Parts
An interactive live film: thirty non-negative data columns, the parts found by the book's multiplicative updates, and the cone they span. Find the fewest parts that rebuild the data to within 2%.
Narration
Here are thirty columns of a non negative matrix, each a point with no negative coordinate, all in this corner of space.
Non negative matrix factorization writes A as W times H, with both factors non negative. The columns of W are parts, and every data column becomes a non negative mix of them: a point inside the cone the parts span.
The rest of the narration is in the book's Companion.
More from Chapter 9
Interactive1:34LoRA: a Low-Rank UpdateChapter 9
Preview1:48LDL Is Completing the SquareChapter 9
Preview1:25Superposition and Sparse CodingChapter 9
Preview1:47Rank-Revealing QR: Pick the Columns in a Good OrderChapter 9
Preview1:56The QR Algorithm: Sweeping a Matrix into Schur FormChapter 9
Preview1:33CUR: Rebuilding a Matrix from Its Own Rows and ColumnsChapter 9