Scalable Force Scheme: a fast method for projecting large datasets

Global dimensionality reduction (DR) methods are widely used to project high-dimensional data into a low-dimensional representation, preserving the overall structure of the dataset. Global nonlinear DR techniques allow one to capture complex features of the data but are limited by their high computational cost, making them an unfeasible choice to process large datasets. Force scheme (FS) is one of the most popular of such examples, being adopted in a wide variety of domains, but limiting its application to small datasets. In this paper, we extend FS to improve its convergence quality and speed, by introducing several concepts from gradient descent (GD) theory and lowering its algorithmic complexity. Our new proposed method is less prone to generate distorted projections due to the presence of artifacts, while significantly improving the running times, allowing for nonlinear global DR projections of large datasets.