Clustering Key Terms, Explained
Getting started with Data Science or need a refresher? Clustering is among the most used tools of Data Scientists. Check out these 10 Clustering-related terms and their concise definitions.
Clustering is a method of data analysis which groups data points together in order to “maximizing the intraclass similarity and minimizing the interclass similarity,” ( by Han, Kamber & Pei) without using predefined labels of points (i.e., an unsupervised learning technique). This post introduces key words for common techniques in cluster analysis.
Different clusters in different color points (From Matthew Mayo‘s Machine Learning Key Terms, Explained)
Feature selection is a data preprocessing step in which redundant and/or irrelevant features are pruned to enhance the quality of clustering. Feature selection can also be integrated directly into the clustering algorithm to gain better locality specific insights.
This is an algorithm is used to estimate the parameters of a specific form assumed of the generative model of data (e.g., mixture of Gaussians).
a) k-means is a method of clustering using distance. It is perhaps the most well-known example of a clustering algorithm. It is the most widely used methods in practical implementations because of its simplicity. The Euclidian distance is used to compute distances. Then, the partitioning representatives correspond to the mean of each cluster.
b) k-medians is quite similar to the k-means method but it uses the median along each dimension, instead of the mean. This approach is more stable to noise and outliers, because the median is usually less sensitive to extreme values in the data.
These methods try to explore the data space at high levels of granularity. Thus, they can be used to reconstruct the entire shape of the data distribution. DBSCAN [1] and STING [2] are two classical examples.
a) The density-based method at any particular point in the data space is defined either in terms of the number of data points in a pre-specified volume of its locality or in terms of a smoother kernel density estimate [3]. This method is naturally defined in a continuous space, therefore, arbitrary data types, e.g. time-series, are not quite as easy to use with density-based methods without specialized transformations.
