Updated September 22, 2023

Introduction to Machine Learning Algorithms

Computers can learn from data and make predictions without explicit programming through machine learning algorithms. They utilize statistical patterns and mathematical models to identify relationships and patterns within datasets, allowing for tasks like classification, regression, clustering, and more. These algorithms play a crucial role in various fields, including artificial intelligence, data science, and predictive analytics, enabling automated decision-making and pattern recognition in diverse applications such as image recognition, natural language processing, and recommendation systems.

Introduction
Importance
Types
Popular Algorithms

Importance of Machine Learning Algorithms

Machine Learning Algorithms are essential because they power the decision-making capabilities of artificial intelligence systems. These algorithms enable machines to learn from data, recognize patterns, make predictions, and automate tasks without explicit programming. Their significance lies in their capacity to extract valuable insights from vast datasets, optimize processes, enhance application personalization, and even contribute to scientific advancements. In fields like healthcare, finance, and autonomous driving, they drive innovation by facilitating predictive modeling and improving efficiency. In essence, Machine Learning Algorithms underpin the transformative potential of AI, making them indispensable for tackling complex problems and shaping the future of technology and industry.

Types of Machine Learning Algorithms

There are various ways to categorize types of machine learning algorithms, but they are typically grouped into classes based on their purpose. The primary categories include the following:

Supervised learning
Unsupervised Learning
Semi-supervised Learning
Reinforcement Learning

What is Supervised Learning

Supervised Learning is where you can consider an instructor who guides the learning. We have a dataset that goes about as an educator, and its job is to prepare the model or the machine. When the model is prepared, it can begin settling on an expectation or choice when new information is given.

Example of Supervised Learning:

You get a lot of photographs with data about what is on them, and after that, you train a model to perceive new photographs.
You have a lot of data about house prices based on their size and location, and you feed it into the model and train it; then, you can predict the price of other houses based on the data you feed.
If you want to predict whether your message is spam or not based on the older message you have, you can predict whether a new message is spam or not.

What is Unsupervised Learning?

The model learns through perception and discovers structures in the information. When the model is given a dataset, it consequently discovers examples and connections in the dataset by making bunches in it. It can’t add marks to the bunch, similar to it can’t state this a gathering of apples or mangoes; however, it will isolate every one of the apples from mangoes.

Assume we displayed pictures of apples, bananas, and mangoes to the model, so it makes bunches and partitions the dataset into those groups in light of certain examples and connections. Presently if another information is bolstered to the model, it adds it to one of the made bunches.

Example of Unsupervised Learning

You have a lot of photographs of 6 individuals yet without data about who is on which one, and you need to isolate this dataset into 6 heaps, each with the photographs of one person.
You have particles, some portion of them are medications, and parts are not; however, you don’t realize which will be which, and you need the calculation to find the medications.

What is Semi-Supervised Learning?

Semi-supervised kind of learning, the calculation is prepared upon a mix of named and unlabeled information. Normally, this blend will contain a limited quantity of named information and a lot of unlabeled information. The fundamental method included is that the software engineer will first group comparable information utilizing an unaided learning calculation and afterward utilize the current named information to name the remainder of the unlabeled information. The ordinary use instances of such kind of calculation have a typical property among them – The obtaining of unlabeled information is generally modest, while naming the said information is over the top expensive. Naturally, one may envision the three kinds of learning calculations as Supervised realizing, where an understudy is under the supervision of an instructor at both home and school, Unsupervised realizing, where an understudy needs to make sense of an idea himself and Semi-Supervised realizing, where an educator shows a couple of ideas in class and gives inquiries as schoolwork which depend on comparable ideas.

Example of Semi-Supervised Learning

It’s outstanding that more information is equal to better quality models in profound learning (up to a specific point of confinement clearly, yet we don’t have that much information more often than not.) Be that as it may, getting marked information is costly. For example, if you need to prepare a model to distinguish winged animals, you can set up a lot of cameras to take pictures of fowl. That is generally modest. Contracting individuals to mark those photos is costly. Consider the possibility that you have an enormous number of pictures of winged animals; however, just contract individuals to mark a small subset of the photos. As it turned out, rather than simply training the models on the marked subset, you can pre-train the model on the whole training set before tweaking it with the named subset, and you show signs of improvement execution along these lines. That is semi-supervised learning. It sets aside your cash.

What is Reinforcement Learning?

It is the capacity of a specialist to collaborate with the earth and discover what the best result is. It pursues the idea of hit and preliminary technique. The operator is remunerated or punished with a point for a right or an off-base answer, and based on the positive reward focuses picked up, the model trains itself. Also, again once prepared, it prepares to foresee the new information introduced to it.

Example of Reinforcement Learning

Displaying ads, according to user like dislikes, optimize for the long-period
Know ads budget used in real-time
inverse reinforcement learning to know customers like dislikes better

Popular Machine Learning Algorithms

There are numerous Machine Learning algorithms in the market currently, and it’s only going to increase considering the amount of research done in this field. Linear and Logistic Regression are generally the first algorithms you learn as a Data Scientist, followed by more advanced algorithms.

Below are some of the Machine Learning algorithms, along with sample code snippets in python:

1. Linear Regression

Linear regression is valuable for discovering the connection between two persistent factors. One is a predictor or autonomous variable; the other is a reaction or ward variable. It searches for measurable relationships, however, not a deterministic relationship. The connection between two factors is said to be deterministic on the off chance that one variable can be precisely communicated by the other. For instance, utilizing temperature in degrees Celsius, it is conceivable to foresee Fahrenheit precisely. A factual relationship isn’t precise in deciding a connection between two factors. For instance, connection somewhere in the range of tallness and weight. The central thought is to get a line that best fits the information. The best-fit line is the one for which the all-out forecast blunder (all information focuses) is as little as expected under the circumstances. The mistake is the separation between the point to the regression line.

As the name suggests, this algorithm could be used in cases where the target variable, which is continuous in nature, is linearly dependent on the dependent variables.

It is represented by:

y = a*x + b + e

where y is the target variable we are trying to predict, a is the intercept, and b is the slope, x is our dependent variable used to make the prediction. This is a Simple Linear Regression, as there is only one independent variable.

In the case of Multiple Linear Regression, the equation would have been:

y = a1*x1 + a2*x2 + …… + a(n)*x(n) + b + e

Here, e is the error term, and a1, a2.. a (n) are the coefficients of the independent variables.

A metric is used to evaluate the model’s performance, which could be Root Mean Square Error, which is the square root of the mean of the sum of the difference between the actual and the predicted values.

The goal of Linear Regression is to find the best-fit line that would minimize the difference between the actual and the predicted data points.

Linear Regression could be written in Python as below:

2. Decision Trees

A Decision tree is a decision-help gadget that uses a tree-like diagram or model of decisions and their potential outcomes, including chance-event results, resource costs, and utility. Explore the image to get a sentiment of what it resembles.

Used for classification and regression problems, the Decision Tree algorithm is one of the most simple and easily interpretable Machine Learning algorithms. Moreover, it is not affected by outliers or missing values in the data and could capture the non-linear relationships between the dependent and the independent variables.

To build a Decision Tree, all features are considered at first, but the feature with the maximum information gain is taken as the final root node based on which the successive splitting is done. This splitting continues on the child node based on the maximum information criteria, and it stops until all the instances have been classified or the data cannot be split further. Decision Trees are often prone to overfitting, and thus it is necessary to tune the hyperparameters like maximum depth, minimum leaf nodes, minimum samples, maximum features, and so on. There is a greedy approach that sets constraints at each step and chooses the best possible criteria for that split to reduce overfitting. There is another better approach called Pruning, where the tree is first built up to a certain pre-defined depth, and then starting from the bottom, the nodes are removed if it doesn’t improve the model.

In sklearn, Decision Trees are coded as:

3. Naive Bayes Classification

Naive Bayes classification is a group of basic probabilistic classifiers dependent on applying Bayes’ theory with strong (unsophisticated) self-governance of the features of Naive Bayes. Some of the certifiable models are:

To stamp an email as spam or not spam.
Order a news story about innovation, governmental issues, or sports.
Check a touch of substance imparting positive emotions or negative sentiments.
Utilized for face acknowledgment programming.

It works on the principle of Bayes Theorem, which finds the probability of an event considering some true conditions.

Bayes Theorem is represented as:

The algorithm is called Naive because it believes all variables are independent and the presence of one variable doesn’t have any relation to the other variables, which is never the case in real life. As a result, naive Bayes could be used in Email Spam classification and in text classification.

Naïve Bayes code in Python:

4. Logistic Regression

Logistic regression is a ground-breaking factual method for demonstrating a binomial result with at least one informative factor. It quantifies the connection between the absolute ward variable and at least one free factor by evaluating probabilities utilizing a logistic capacity, the combined logistic appropriation.

In terms of maintaining a linear relationship, it is the same as Linear Regression. However, unlike in Linear Regression, the target variable in Logistic Regression is categorical, i.e., binary, multinomial, or ordinal in nature. Moreover, the choice of the activation function is important in Logistic Regression as for binary classification problems, the log of odds in favor, i.e., the sigmoid function, is used.

In the case of a multi-class problem, the softmax function is preferred as a sigmoid function takes a lot of computation time.

The metric used to evaluate a classification problem is generally Accuracy or the ROC curve. The more the area under the ROC, the better the model. For example, a random graph would have an AUC of 0.5. A value of 1 indicates the most accuracy, whereas 0 indicates the least accuracy.

Logistic Regression could be written in learning as:

5. Ordinary Least Squares Regression

The least squares is a strategy for performing direct regression. direct regression is the undertaking of fitting a line through a lot of focus. There are various potential procedures to do this, and the “ordinary least squares” system goes like this— You can draw a line, and after that, for all of the data centers, measure the vertical detachment between the point and the line, and incorporate these up; the fitted line would be the place this aggregate of partitions is as meager as could be normal in light of the current situation.

6. Clustering

Clustering is a significant idea with regard to unaided learning. It manages to find a structure or example in a gathering of uncategorized information for the most part. Clustering calculations will process your information and discover characteristic clusters(groups) in the event that they exist in the information. You can likewise alter what number of bunches your calculations ought to distinguish. It enables you to alter the granularity of these gatherings.

There are various kinds of clustering you can use

Selective (apportioning)
Model: K-means
Agglomerative
Model: Hierarchical clustering
Covering
Model: Fuzzy C-Means
Probabilistic

Clustering algorithm Types

Hierarchical clustering
K-means clustering
K-NN (k nearest neighbors)
Principal Component Analysis
Solitary Value Decomposition
Independent Component Analysis

1. Hierarchical Clustering

Hierarchical clustering is a calculation that constructs a pecking order of groups. It starts with every one of the information which is doled out to their very own bunch. Here, two close groups will be in a similar bunch. This calculation closes when there is just one group left.

2. K-means Clustering

K means it is an iterative clustering calculation that encourages you to locate the most noteworthy incentive for each emphasis. At first, the ideal number of groups is chosen. In this clustering technique, you have to bunch the information that focuses on k gatherings. A bigger k means littler gatherings with greater granularity similarly. A lower k means bigger gatherings with less granularity.

The yield of the calculation is a gathering of “names.” It allows information to point to one of the k gatherings. In k-means clustering, each gathering is characterized by making a centroid for each gathering. The centroids are like the core of the bunch, which catches the focuses nearest to them and adds them to the group.

K-mean clustering further characterizes two subgroups.

Agglomerative clustering: This sort of K-means clustering begins with a fixed number of bunches. Then, it designates all information into an accurate number of groups. This clustering strategy doesn’t require the number of groups K as info. The agglomeration procedure begins by shaping every datum as a solitary bunch. This strategy utilizes some separation measures, lessens the number of bunches (one in every emphasis) by combining processes. In conclusion, we have one major group that contains every one of the articles.
Dendrogram: In the Dendrogram clustering technique, each level will speak to a conceivable bunch. The tallness of the dendrogram demonstrates the degree of similitude between two join bunches. The closer to the base of the procedure, the progressively comparable bunch, is the finding of the gathering from dendrogram, which isn’t characteristic and, for the most part, abstract.

So far, we have worked with supervised learning problems where there is a corresponding output for every input. Now, we would learn about unsupervised learning, where the data is unlabelled and needs to be clustered into specific groups. There are several clustering techniques available. However, the most common of them is the K-means clustering. Ink-means, k refers to the number of clusters that need to be set in prior to maintaining maximum variance in the dataset. Once the k is set, the centroids are initialized. The centroids are then adjusted repeatedly so that the distance between the data points within a centroid is maximum and the distance between two separate is maximum. Euclidean distance, Manhattan distance, etc., are some of the distance formula used for this purpose.

The value of k could be found from the elbow method.

K-means clustering is used in e-commerce industries where customers are grouped together based on their behavioral patterns. It could also be used in Risk Analytics.

Below is the python code:

K-Nearest neighbors

K-nearest neighbor is the most straightforward of all AI classifiers. However, it varies from other AI procedures in that it doesn’t deliver a model. Instead, it is a straightforward calculation that stores every single accessible case and characterizes new examples dependent on a likeness measure.

It works very well when there is a separation between models. However, the learning rate is moderate when the preparation set is enormous, and the separation figuring is nontrivial.

Machine Learning Algorithms could be used for both classification and regression problems. The idea behind the KNN method is that it predicts the value of a new data point based on its K Nearest Neighbors. K is generally preferred as an odd number to avoid any conflict. While classifying any new data point, the class with the highest mode within the Neighbors is taken into consideration. While for the regression problem, the mean is considered as the value.

I learned the KNN is written as:

KNN is used in building a recommendation engine.

Principal Components Analysis

On the off chance that you need a higher-dimensional space. You have to choose a reason for that space and just the 200 most significant scores of that premise. This base is known as a principal component. The subset you select comprise is another space that is little in size contrasted with unique space. It keeps up, however, much of the multifaceted nature of information as could be expected.

7. Support Vector Machines

A classification algorithm where a hyperplane separates the two classes. In a binary classification problem, two vectors from two distinct classes are considered known as the support vectors, and the hyperplane is drawn at a maximum distance from the support vectors.

As you can see, a single line separates the two classes. However, in most cases, the data would not be perfect, and a simple hyperplane would not be able to separate the classes. Hence, you need to tune parameters such as Regularization, Kernel, Gamma, and so on.

The kernel could be linear or polynomial, depending on how the data is separated. In this case, the kernel is linear in nature. In the case of Regularization, you need to choose an optimum value of C, as the high value could lead to overfitting while a small value could underfit the model. Gamma defines the influence of a single training example. Points close to the line are considered in high gamma and vice versa for low gamma.

In sklearn, SVM is written as:

8. Random Forest

To reduce overfitting in the Decision Tree, it is required to reduce the variance of the model, and thus the concept of bagging came into place. Bagging is a technique where the output of several classifiers is taken to form the final output. Random Forest is one such bagging method where the dataset is sampled into multiple datasets, and the features are selected at random for each set. Then on each sampled data, the Decision Tree algorithm is applied to get the output from each mode. In the case of a Regression problem, the mean of the output of all the models is taken, whereas, in the case of classification problems, the class which gets the maximum vote is considered to classify the data point. Random Forest is not influenced by outliers, missing values in the data, and it also helps in dimensionality reduction as well. However, it is not interpretable, which is a drawback for Random Forest.

In Python, you could code Random Forest as:

Conclusion

Data Scientist is the sexiest job in the 21^st century, and Machine Learning is certainly one of its key areas of expertise. To be a Data Scientist, one needs to possess an in-depth understanding of all these algorithms and also several other new techniques such as Deep Learning.

Quiz Result
Total Questions	Correct Answers	Wrong Answers	Percentage