A Dive into Un-Supervised Machine Learning Algorithms

When we are presented with un-labeled data, un-supervised learning is used to extract underlying patterns from it and no supervision or intervention from humans is done. Un-supervised machine learning works with a large amount of data that is often un-structured and contains a lot of noise.

There are three main types of un-supervised machine learning, clustering, association, and dimensional reduction.

Clustering refers to categorizing the items in the dataset in different classes based on their similarities and differences. 

Association refers to finding association between the items of datasets and making decisions based on these relationships.

Dimensional reduction involves removing the redundant variables and features from the dataset while preserving the important ones to make data analysis and visualization easier.

Algorithms

Different algorithms are applied in different scenarios to extract patterns from the data and analyze them. These algorithms work without explicit supervision or labeled outcomes.

• Neural Networks

Neural networks, also known as deep learning, are the type of algorithms that are inspired by the human brain. They involve different nodes, called neurons arranged in layers and rely on the transmission of data from neurons in one layer to the next. Data is input in the first layer and the output of that layer is fed into the next layer as input and so on. There is an input layer (that takes in the raw data), output layer (that gives the final output) and hidden layers (which processes the data and passes it on, these can be many in number). 

Neural networks can work on unstructured data and extract patterns from it requiring little to no human intervention. They are even capable of learning from mistakes and produce outputs with great accuracy. Neural networks are categorized in several types based on their architecture and connectivity patterns.

1. Feed Forward Networks (FNN) – the data travels in one direction i.e. from the input node, through the hidden layers, towards the output node. Every node of a layer is connected to every other node in the next layer.
2. Recurrent Neural Networks (RNN) – the data travels in a cyclic path which makes them dynamic and suited to handle sequential data. The recurrent structure can retain previous outputs and combined with recent outputs, makes decisions based on them.
3. Convolutional Neural Networks (CNN) – grid like data is better suited to CNNs as each hidden layer specializes in processing a different feature of it and generates the best result. This is why it is best for image recognition.
4. Auto Encoders – high dimensional data is compressed into a lower dimension and then reconstructed to form the input from it. This is why these are best for dimension reduction, image classification etc.

• Principal Component Analysis

Principal component analysis is a dimension reduction technique where a large dataset is broken down and compressed into smaller datasets, while preserving as many necessary features as possible. The smaller version makes it easier to work with and analyze. The goal is to identify the variance and capture it in the form of principal components. These principles are independent of one another and contain a different aspect of variability.

• Apriori Algorithm

Apriori algorithm is a very reliable association rule-based algorithm that aims to find the relationships of different variables in the dataset to generate output. It is most frequently used in market-based analysis, and databases that contain transactions. It works with itemset (those items whose support is greater than the threshold than their own individually) based on support and confidence.  Support means the frequency with which the itemset appears in the dataset and confidence is the reliability of the association rule. The property that makes this algorithm efficient is that any subset of a frequent itemset should also be frequent.

• K-Means Clustering

k-means clustering is a clustering technique that divides the raw and unlabeled input data into k number of clusters based on similar features or attributes. The number of clusters, that is K should be predefined. It aims to minimize the distance between the center point (centroid) and the data points in the same clusters.

• Anomaly Detection

Anomaly detection, also known as the outlier detection method, is a technique in unsupervised machine learning that is employed to detect abnormal behavior or data points that deviate from normal data points. These abnormal data points can be outliers (short changes), sudden event change, or drifts (slow change). Anomaly detection is great to detect any fraud in transaction or a cybersecurity threat. Detection of anomalies can lead to immediate investigation and prevention of a loss or breach of security. 

• Singular Value Decomposition

Singular value decomposition is a dimensional reduction method used with principal component analysis and decomposes a matrix into three product matrices. The smaller matrices are easier to work with as they are free of redundant data points. The smaller matrices are the subsets of the large matrix of data being input and it has all the values necessary for approximation of the original matrix with lower dimensionality.

Remarks

These un-supervised learning algorithms have a vast variety of applications and have their own pros and cons. These are employed based on the data being provided with and the desirable outputs. The goal is to produce the most accurate outcome out of the real time data so that appropriate decisions can be made and certain risks can be mitigated.