An alternative for data visualization using space-filling curve

Dimensionality reduction helps data analysts and machine learning designers to visualize in low dimension structures lying in high dimension. This is a basic but crucial operation, to discover relationship between variables, considering the difficulties to tweek machine learning algorithm. The data have not to be consider as a black-box but can be visualized, leading to better decision making. Inspired from previous works, this article proposes to create a dimensionality reduction method based on space-filling curves (SFCs). Of course, the Hilbert curve was considered (guided by reflected binary gray code pattern) but also alternative high locality SFCs, recently identified. Mapping algorithms working with alternative curves are provided, and illustrated through a numerical example. Mapping a D-dimensional point to a 1 D index is usual but developing an algorithm for reverse mapping, i.e. from 1 D index to 2 D or 3 D point is more original and can allow the visualization of data. The work position is specified and justifications are given. A discussion on the choice of parameters (order of curves n and n' ) is led in order to guide the user to select good parameters (to define a bijection between original data space and projected space). Experiments are conducted to compare our proposition to state of the art approaches (PCA, MDS, t-SNE, UMAP) over seven dataset involving from 3 D to 16 D and covering diverse topologies. The results show interesting ability on data visualization. Compare to standard techniques, the time computing is low, which is an interesting property in regards to the amount of data today created.