Analyzing high-dimensional data poses many challenges due to the "curse of dimensionality". Not all high-dimensional data exhibit these characteristics because many data sets have correlations, which led to the notion of intrinsic dimensionality. Intrinsic dimensionality describes the local behavior of data on a low-dimensional manifold within the higher dimensional space. We discuss this effect, and describe a surprisingly simple approach modification that allows us to reduce local intrinsic dimensionality of individual points. While this unlikely will be able to "cure" all problems associated with high dimensionality, we show the theoretical impact on idealized distributions and how to practically incorporate it into new, more robust, algorithms. To demonstrate the effect of this adjustment, we introduce the novel Intrinsic Stochastic Outlier Score (ISOS), and we propose modifications of the popular t-Stochastic Neighbor Embedding (t-SNE) visualization technique for intrinsic dimensionality, intrinsic t-Stochastic Neighbor Embedding (it-SNE).
展开▼
