Similarity measure

Use in clustering

Clustering or Cluster analysis is a data mining technique that is used to discover patterns in data by grouping similar objects together. It involves partitioning a set of data points into groups or clusters based on their similarities. One of the fundamental aspects of clustering is how to measure similarity between data points.

Similarity measures play a crucial role in many clustering techniques, as they are used to determine how closely related two data points are and whether they should be grouped together in the same cluster. A similarity measure can take many different forms depending on the type of data being clustered and the specific problem being solved.

One of the most commonly used similarity measures is the Euclidean distance, which is used in many clustering techniques including K-means clustering and Hierarchical clustering. The Euclidean distance is a measure of the straight-line distance between two points in a high-dimensional space. It is calculated as the square root of the sum of the squared differences between the corresponding coordinates of the two points. For example, if we have two data points ${\displaystyle (x_1,y_1)}$ and ${\displaystyle (x_2,y_2)}$, the Euclidean distance between them is${\displaystyle d=\surd [(x_2-x_1{)}^2+(y_2-y_1{)}^2]}$.

Another commonly used similarity measure is the Jaccard index or Jaccard similarity, which is used in clustering techniques that work with binary data such as presence/absence data or Boolean data; The Jaccard similarity is particularly useful for clustering techniques that work with text data, where it can be used to identify clusters of similar documents based on their shared features or keywords. It is calculated as the size of the intersection of two sets divided by the size of the union of the two sets: ${\displaystyle J(A,B)=\frac{A\bigcap B}{A\bigcup B}}$.

Similarities among 162 Relevant Nuclear Profile are tested using the Jaccard Similarity measure (see figure with heatmap). The Jaccard similarity of the nuclear profile ranges from 0 to 1, with 0 indicating no similarity between the two sets and 1 indicating perfect similarity with the aim of clustering the most similar nuclear profile.

Manhattan distance, also known as Taxicab geometry, is a commonly used similarity measure in clustering techniques that work with continuous data. It is a measure of the distance between two data points in a high-dimensional space, calculated as the sum of the absolute differences between the corresponding coordinates of the two points ${\displaystyle \left|x_1-x_2\right|+\left|y_1-y_2\right|}$.

When dealing with mixed-type data, including nominal, ordinal, and numerical attributes per object, Gower's distance (or similarity) is a common choice as it can handle different types of variables implicitly. It first computes similarities between the pair of variables in each object, and then combines those similarities to a single weighted average per object-pair. As such, for two objects ${\displaystyle i}$ and ${\displaystyle j}$ having ${\displaystyle p}$ descriptors, the similarity ${\displaystyle S}$ is defined as:$${\displaystyle S_{ij}=\frac{\sum _{k=1}^p{w}_{ijk}{s}_{ijk}}{\sum _{k=1}^p{w}_{ijk}},}$$ where the ${\displaystyle w_{ijk}}$ are non-negative weights and ${\displaystyle s_{ijk}}$ is the similarity between the two objects regarding their ${\displaystyle k}$-th variable.

In spectral clustering, a similarity, or affinity, measure is used to transform data to overcome difficulties related to lack of convexity in the shape of the data distribution. The measure gives rise to an ${\displaystyle (n,n)}$-sized similarity matrix for a set of n points, where the entry ${\displaystyle (i,j)}$ in the matrix can be simply the (reciprocal of the) Euclidean distance between ${\displaystyle i}$ and ${\displaystyle j}$, or it can be a more complex measure of distance such as the Gaussian ${\displaystyle e^{-\Vert s_1-s_2{\Vert }^2/2{\sigma }^2}}$. Further modifying this result with network analysis techniques is also common.

The choice of similarity measure depends on the type of data being clustered and the specific problem being solved. For example, working with continuous data such as gene expression data, the Euclidean distance or cosine similarity may be appropriate. If working with binary data such as the presence of a genomic loci in a nuclear profile, the Jaccard index may be more appropriate. Lastly, working with data that is arranged in a grid or lattice structure, such as image or signal processing data, the Manhattan distance is particularly useful for the clustering.

Use in recommender systems

Similarity measures are used to develop recommender systems. It observes a user's perception and liking of multiple items. On recommender systems, the method is using a distance calculation such as Euclidean Distance or Cosine Similarity to generate a similarity matrix with values representing the similarity of any pair of targets. Then, by analyzing and comparing the values in the matrix, it is possible to match two targets to a user's preference or link users based on their marks. In this system, it is relevant to observe the value itself and the absolute distance between two values. Gathering this data can indicate a mark's likeliness to a user as well as how mutually closely two marks are either rejected or accepted. It is possible then to recommend to a user targets with high similarity to the user's likes.

Recommender systems are observed in multiple online entertainment platforms, in social media and streaming websites. The logic for the construction of this systems is based on similarity measures.

Use in sequence alignment

Similarity matrices are used in sequence alignment. Higher scores are given to more-similar characters, and lower or negative scores for dissimilar characters.

Nucleotide similarity matrices are used to align nucleic acid sequences. Because there are only four nucleotides commonly found in DNA (Adenine (A), Cytosine (C), Guanine (G) and Thymine (T)), nucleotide similarity matrices are much simpler than protein similarity matrices. For example, a simple matrix will assign identical bases a score of +1 and non-identical bases a score of −1. A more complicated matrix would give a higher score to transitions (changes from a pyrimidine such as C or T to another pyrimidine, or from a purine such as A or G to another purine) than to transversions (from a pyrimidine to a purine or vice versa). The match/mismatch ratio of the matrix sets the target evolutionary distance. The +1/−3 DNA matrix used by BLASTN is best suited for finding matches between sequences that are 99% identical; a +1/−1 (or +4/−4) matrix is much more suited to sequences with about 70% similarity. Matrices for lower similarity sequences require longer sequence alignments.

Amino acid similarity matrices are more complicated, because there are 20 amino acids coded for by the genetic code, and so a larger number of possible substitutions. Therefore, the similarity matrix for amino acids contains 400 entries (although it is usually symmetric). The first approach scored all amino acid changes equally. A later refinement was to determine amino acid similarities based on how many base changes were required to change a codon to code for that amino acid. This model is better, but it doesn't take into account the selective pressure of amino acid changes. Better models took into account the chemical properties of amino acids.

One approach has been to empirically generate the similarity matrices. The Dayhoff method used phylogenetic trees and sequences taken from species on the tree. This approach has given rise to the PAM series of matrices. PAM matrices are labelled based on how many nucleotide changes have occurred, per 100 amino acids. While the PAM matrices benefit from having a well understood evolutionary model, they are most useful at short evolutionary distances (PAM10–PAM120). At long evolutionary distances, for example PAM250 or 20% identity, it has been shown that the BLOSUM matrices are much more effective.

The BLOSUM series were generated by comparing a number of divergent sequences. The BLOSUM series are labeled based on how much entropy remains unmutated between all sequences, so a lower BLOSUM number corresponds to a higher PAM number.

Use in computer vision

The most common method for comparing two images in content-based image retrieval (typically an example image and an image from the database) is using an image distance measure. An image distance measure compares the similarity of two images in various dimensions such as color, texture, shape, and others. For example, a distance of 0 signifies an exact match with the query, with respect to the dimensions that were considered. As one may intuitively gather, a value greater than 0 indicates various degrees of similarities between the images. Search results then can be sorted based on their distance to the queried image. Many measures of image distance (Similarity Models) have been developed.