Embedding Method for Structural Preservation via Pairwise Attractiveness

This study proposes a dimensionality reduction algorithm that projects high-dimensional data into a lower dimensional space while preserving directional and density-related characteristics. In conventional dimensionality reduction algorithms, the loss function was defined by some distance or similarity between the data, or by the local distance using an exponential function. This study allows the weights of cosine similarity and Euclidean distance to be freely defined and then defines a loss function that optimally separates clusters in the low-dimensional space after dimensionality reduction, thereby providing a dimensionality reduction algorithm that not only improves the accuracy of clustering but also makes it easy to consider data in highdimensional spaces.