Clustering is done to simplify data management so that intelligent models can distinguish different information from each other. When dealing with a small set of attributes, categorizing is easy.

For example, if we have blue, black, red, and green pens, it is possible to classify them into four groups. Still, other features in the same set, such as size, manufacturer, weight, price, and other attributes, Get a little complicated.

Now suppose we want to categorize in a collection of thousands of records and hundreds of attributes; what strategy should we use to categorize correctly?

What is clustering?

Segmentation, the more complete and accurate term for cluster analysis, refers to the process by which a set of objects can be assigned to separate groups. The members of each cluster are similar to each other based on their characteristics, and in contrast, the degree of similarity between the clusters is low.

Clustering is done with the aim of labeling objects so that it is easy to identify objects that are members of different groups. In this method, the data is divided into meaningful groups whose contents of each cluster have similar properties but different from other objects that are placed in other groups.

The clustering mechanism is used in large data sets and cases where the number of data properties is large.

As mentioned, in the clustering process, groupings of objects are grouped based on similar properties. These are exploratory data mining algorithms that use conventional statistical analysis methods to access objects.

The main difference between clustering analysis and classification analysis is the lack of primary tags for observations. In clustering based on common features and methods of measuring distance or similarity between objects, tags are assigned automatically. And in classification, there are early tags, and predictive algorithms must be used to tag new observations.

In data mining, the result of groups is considered, and in automatic classification, diagnostic power is considered. Depending on the different methods of measuring similarities or clustering algorithms, it is likely that the clustering results will be different for the fixed data set.

Cluster terms include groups with short distances between cluster members, dense areas of data space, distances, or specific statistical distributions. Therefore, clustering can be done as a multi-objective optimization problem.

Proper clustering algorithms and parametric settings (such as distance vector function, density threshold, or expected number of clusters) depend on the individual set of data set and the specific use of the results.

Cluster analysis is not an automated method but an iterative process of knowledge discovery or multi-objective interactive optimization based on trial and error. In most cases, the preprocessed data and model parameters must be modified to achieve the desired result.

It isn’t easy to clearly define the concept of clustering, one of the reasons being that there are many clustering algorithms, yet they all have one thing in common that has at least one set of data objects. Researchers use different cluster models and have designed different algorithms for each of these cluster models.

Clustering models

Most algorithms and clustering methods indeed have the same structure. However, there are differences in the way similarities or distances are measured and the label selection for the objects of each cluster in these methods. Accordingly, their classification provides a better view of the clustering method used in the algorithms.

In general, clustering algorithms can be divided into 4 main groups of discrete, hierarchical, density-based, and model-based clustering algorithms, although some sources offer more models as follows.

Connected models: For example, hierarchical clustering creates models based on connected distances.

Central models: For example, the k-means algorithm shows each cluster with an average vector.

Distribution Models: Clusters use statistical distributions, such as the normal multivariate distribution used in the maximum expectation algorithm.

Density models: DBSCAN and OPTICS define clusters as densely connected areas in the data space.

Subspace models: Known as common clusters or two-state clusters. Clusters are modeled with both cluster members and related properties.

Group models: Some algorithms do not provide a modified model for their results and only provide grouping information.

Graph-based models: A class, i.e., a subset of nodes in a graph so that both nodes in the subset are connected by an edge that can be considered as a primary form of the cluster.

Neural models: The uncontrolled neural network is the most well-known self-organizational map, and usually, these models can be similar to one or more of the above models, including subspace models, when neural networks are a form of principal component analysis or independent element analysis. Is.

Clustering is basically a set of clusters that usually include all the objects in the data set. In addition, the relationship of clusters to each other can be defined, for example, the hierarchy of clusters embedded in each other.

Classification based on difficulty

Clustering can be characterized by the difficulty of differentiation as follows:

Strict clustering: Every object belongs to the cluster or not.

Soft clustering (also: fuzzy clustering): Each object belongs to a certain degree of each cluster (e.g., probability of cluster dependence)

Precise Separation Clustering (Partitioning): Each object belongs to exactly one cluster.

Accurate Separation Clustering with Discontinuity: Objects can not belong to any cluster.

Overlapping clustering (also: alternate clustering, multiple clustering): Objects may belong to more than one cluster; Usually includes hard clusters

Hierarchical clustering: Objects that belong to the child cluster also belong to the parent cluster.

Subspace clustering: While overlapping clustering, defined in a unique subspace, clusters are not expected to overlap.

The best clustering algorithm for a particular problem often has to be chosen experimentally unless there is a mathematical reason for preferring one clustering model.

It should be noted that an algorithm that is designed for one type of model fails in a data set that contains the fundamental difference of the model. For example, k-means cannot find non-convex clusters.

Hierarchical clustering

Connection-based clustering, known as hierarchical clustering, is based on the basic idea of ​​objects that are more closely related to nearby objects than to more distant ones. These algorithms connect objects to create clusters based on their distance.

The cluster can generally be described by the maximum distance required to connect the cluster components. At different distances, different clusters are formed, which can be represented using a dendrogram, which explains what the common name for “hierarchical clustering” comes from: These algorithms do not provide a data set partition but an extensive hierarchy.

Provides clusters that merge at regular intervals. Hierarchical clustering includes two types of clustering:

single linkage -1 This method, also known as Bottom-Up and Agglomerative methods, is a method in which each data is first considered a cluster. Then, using an algorithm, each time clusters with close properties are merged, which continues to reach several separate clusters.

The problem with this method is that it is sensitive to noise and consumes a lot of memory.

Complete linkage -2 In this method, known as the Top-Down and Divisive method, all data is considered a cluster. Using an iterative algorithm, each time the data has the least similarity with other data is divided into separate clusters.

Is divided. This continues until one or more clusters of a member are created. The noise problem is solved in this way.