Clustering high dimensional data by Ira Assent. (Assent, I. (2012), Clustering high dimensional data. WIREs Data Mining Knowl Discov, 2: 340–350. doi: 10.1002/widm.1062)
High-dimensional data, i.e., data described by a large number of attributes, pose specific challenges to clustering. The so-called ‘curse of dimensionality’, coined originally to describe the general increase in complexity of various computational problems as dimensionality increases, is known to render traditional clustering algorithms ineffective. The curse of dimensionality, among other effects, means that with increasing number of dimensions, a loss of meaningful differentiation between similar and dissimilar objects is observed. As high-dimensional objects appear almost alike, new approaches for clustering are required. Consequently, recent research has focused on developing techniques and clustering algorithms specifically for high-dimensional data. Still, open research issues remain. Clustering is a data mining task devoted to the automatic grouping of data based on mutual similarity. Each cluster groups objects that are similar to one another, whereas dissimilar objects are assigned to different clusters, possibly separating out noise. In this manner, clusters describe the data structure in an unsupervised manner, i.e., without the need for class labels. A number of clustering paradigms exist that provide different cluster models and different algorithmic approaches for cluster detection. Common to all approaches is the fact that they require some underlying assessment of similarity between data objects. In this article, we provide an overview of the effects of high-dimensional spaces, and their implications for different clustering paradigms. We review models and algorithms that address clustering in high dimensions, with pointers to the literature, and sketch open research issues. We conclude with a summary of the state of the art.
The author has a clever example (figure 4) of why adding dimensions can decrease the discernment of distinct groups in data. A problem that worsens as the number of dimensions increases.
Or does it? Or is it the case that by weighting all dimensions equally we get the result we deserve?
My counter-example would be introducing you to twin sisters. As the number of dimensions increased, so would the similarity that would befoul any clustering algorithm.
But the important dimension, their names, is sufficient to cluster attributes around the appropriate data points.
Is the “curse of dimensionality” rather a “failure to choose dimensions wisely?”