ABSTRACT
While many dimensionality reduction algorithms preserve the local structure of the data in low dimensional space, they do not preserve both the local and global geometry in a single mapping. However, a relatively new dimensionality reduction technique called t-distributed Stochastic Neighbor Embedding (t-SNE) captures the local structure as well as reveals the global structure at different scales. This technique was proposed as a variant of Stochastic Neighbor Embedding (SNE) to overcome the limitations of SNE by making the cost function symmetric, and it uses t-distribution instead of Gaussian distribution. This chapter starts with a brief discussion of SNE and its limitations followed by a detailed explanation of t-SNE and how t-SNE finds low dimensional representation of data by overcoming the limitations of SNE. Further, the advantages and shortcomings of this algorithm are also discussed. Finally, the chapter ends with examples using datasets along with a tutorial to better understand the working of this dimensionality reduction technique.
