Data Science and Machine Learning: Making Data-Driven Decisions

1 Case Study

Python, for Data Scientists and Machine Learning specialists, is a lingua franca owing to the immense promise of this widely-used programming language. To strengthen your Python foundations, this module focuses on NumPy, Pandas, and Data Visualization.

• Numpy
  Numpy is a Python package for scientific computing that enables one to work with multi-dimensional arrays and matrices.
• Pandas
  Pandas is an open-source and powerful library in Python that is used to analyze and manipulate data.
• Data Visualization
  Data Visualization means dealing with the graphic representation of data, which effectively generates insights from data by using matplotlib, seaborn, etc., libraries.

1 Case Study

This week’s chapter will help you understand the role of statistics in helping organizations make effective decisions, learn its most widely-used tools, and learn to solve business problems using analysis, data interpretation, and experiments. It will cover the following topics:

• Descriptive Statistics
  It gives you the fundamental measures of a statistical summary of the data.
• Inferential Statistics
  It will explore the areas of distributions and parameter estimation, ultimately allowing you to make inferences from the data.

Here, you will learn about one of the essential aspects of ML - Unsupervised Learning.

• What is unsupervised learning, and why is it challenging?
  Unsupervised learning algorithms will assist you in analyzing and clustering unlabelled data sets. This chapter teaches you about unsupervised learning and the challenges encountered using these algorithms.
• Examples of unsupervised learning
  This chapter will make you understand the implementation of several unsupervised learning algorithms with examples.

2 Case Studies

Clustering is an unsupervised learning technique to group similar sets of data points. The next module of the course in Data Science from MIT will introduce you to the widely used clustering techniques, i.e., K-means clustering.

• What is clustering?
  Here, we will discuss the basic intuition behind clustering and why it is incredibly prevalent in numerous industries.
• When to use clustering?
  This chapter will teach you the procedure for using clustering techniques.
• K-means preliminaries
  This chapter will make you understand a few preliminaries before beginning with K-means clustering.
• The K-means algorithm
  The K-means algorithm in clustering is one of the most commonly implemented unsupervised learning algorithms for resolving clustering problems in Data Science or Machine Learning.
• How to evaluate clustering?
  This chapter will make you familiar with the procedure to evaluate clustering.
• Beyond K-means: What makes a cluster?
  Here, you will learn several techniques that make a cluster.
• Beyond K-means: Other notions of distance
  This chapter will familiarize you with several other types of distance methods in clustering and teach you the use cases for the same.
• Beyond K-means: Data and pre-processing
  Data pre-processing is a technique to clean raw data to use for machine learning applications. It is the initial and foremost step when implementing a machine learning project.
  This chapter will discuss why pre-processing is necessary for Data Science and all the steps involved.
• Beyond clustering
  In this chapter, you will understand all the crucial topics beyond clustering and its applications.

2 Case Studies

Spectral clustering is one of the most widely implemented techniques for cluster graphs and networks. Here, you will learn about spectral, modularity clustering, and the PCA algorithm. This module will discuss spectral clustering and its components and embeddings.

• What if we do not have features to describe the data or not all are meaningful?
  This chapter will teach you how to provide a solution if you do not have any features to describe the data or if not all are meaningful.
• Finding the principal components in data and applications
  Principal Component Analysis is a method to reduce the complexity of an unsupervised machine learning model. In Layman’s terms, Principal Component Analysis is like eliminating the input variables for a predictive model to avoid overfitting.
• The magic of eigenvectors I
  In this chapter, you will understand the procedure to implement eigenvectors in a matrix.
• Clustering in graphs and networks
  Here, you will gain an understanding of clustering in graphs and networks.
• Features from graphs: The magic of eigenvectors II
  Here, you will understand the procedure to implement eigenvectors in a matrix using several features from graphs.
• Spectral clustering
  Spectral clustering will enable you to reduce complex multi-dimensional datasets into identical data clusters in rarer dimensions.
• Modularity Clustering
  The measure of the strength of a network division into clusters is called Modularity clustering.
• Embeddings: New features and their meaning
  An embedding is a moderately low-dimensional space to translate high-dimensional vectors, which assists in making it easier to do machine learning on enormous inputs.

2 Case Studies

Here, you will learn about linear and nonlinear regression together with their extensions, including the crucial case of logistic regression for binary classification and causal inference, where the goal is to understand the effects of actively manipulating a variable as opposed to passively measuring it.

• Linear regression with one and several variables
  Here, you will understand the procedure to implement linear regression with one and several variables.
• Linear regression for prediction
  This chapter will familiarize you with the procedure to implement linear regression for predictive analysis.
• Linear regression for causal inference
  This chapter will familiarize you with the procedure to implement linear regression for causal inference.
• Logistic and other types of nonlinear regression
  Logistic regression is a simple classification algorithm in Machine Learning that predicts the categorical dependent variables using independent variables.
  This chapter will familiarize you with all the fundamentals of Logistic Regression and other types of nonlinear regression in Machine Learning.​​​​​​​

1 Case Study

In the next module of this Data Science for working professionals course, you will learn about modern regression with high-dimensional data or finding a needle in a haystack. For large datasets, it becomes necessary to sort out which variables are relevant for prediction and which are not. Recent years have witnessed the development of new statistical techniques, such as Lasso or Random Forests, that are computationally superior to large datasets and automatically select relevant data.

• Making good predictions with high-dimensional data
  This chapter will teach you the process of making good predictions with high-dimensional data.
• Avoiding overfitting by validation and cross-validation
  Overfitting occurs when a model over-trains the data. In Layman's terms, suppose a model learns the detail and noise within the training data. In that case, the training data will negatively affect the performance of the model on new data.
  This chapter will teach you the process of avoiding overfitting through validation and cross-validation techniques.
• Regularization by Lasso, Ridge, and their modification
  Here, you will understand regularization by Lasso, Ridge, and their modification.
• Regression Trees, Random Forest, Boosted Trees
  Regression Trees are built using binary recursive partitioning, an iterative process that splits the data into partitions or branches. It later splits each portion into smaller groups as the process advances every branch.
  Random Forest is a prevalent supervised Machine Learning algorithm that constitutes numerous decision trees on the given innumerable subsets of a dataset. Later, it will calculate the average to enhance the data set's predictive accuracy.Boosting is a meta-algorithm in Machine Learning, which transforms robust classifiers from several weak classifiers.
  Boosting can be distinguished as Gradient boosting and Adaptive (ADA) boosting.

2 Case Studies

This part will cover regression and causal inference to explain why “correlation does not imply causation” and how we can overcome this intrinsic limitation of regression by resorting to randomized control studies or controlling for confounding.

• Randomized Control Trials
  This chapter will teach you the process of identifying and working with Randomized Control Trials.
• Observational Studies with Confounding
  Confounding is a common hazard of observational clinical research opposing randomized experiments. Yet, it can easily pass unrecognized, although its recognition is essential for significantly interpreting causal relationships, like evaluating treatment effects.

1 Case Study

In this module of the MIT Data Science certificate program, you will learn Hypothesis testing and several classification algorithms. Hypothesis Testing is a technique to perform experiments using the observed/surveyed data. As the name indicates, classification is a technique to classify a data set into different categories and can be performed on both structured and unstructured data.

• What are anomalies? What is fraud? Spams?
  Anomalies occur when databases are inadequately planned and un-normalized, where all the data is stored in one table. Fraud, as the name suggests, is a fraudulent act with no authorization. Spam is unsolicited digital communication, such as sending messages, emails, etc., to vast amounts of people for commercial purposes.
  In this chapter, you will understand the procedure to detect anomalies, fraud, and filter spam in Machine Learning.
• Binary Classification: False Positive/Negative, Precision / Recall, F1-Score
  Binary classification is a supervised machine learning technique, where the categories are predefined and classified into new probabilistic observations. When there are two categories, it is called binary classification.
• Logistic and Probit regression: Statistical binary classification
  Probit regression is a method where the dependent variable takes only two values. This chapter will discuss all the essential concepts, like Logistic regression, Probit regression, and Statistical binary classification.
• Hypothesis testing: Ratio Test and Neyman-Pearson p-values: Confidence
  Here, you will gain an understanding of all the critical concepts of hypothesis testing.
• Support vector machine: Non-statistical classifier
  Support Vector Machine, shortened to SVM, is another popular Machine Learning algorithm used for regression and classification problems.
• Perceptron: Simple classifier with elegant interpretation
  A perceptron is an artificial neuron, or plainly, a mathematical model of a biological neuron. This chapter will familiarize you with perceptron and its various concepts.

1 Case Study

Here, the learners will understand all the critical concepts of Deep Learning, such as image classification, back-propagation, transfer learning, NLP, speech recognition, and much more.

• What is image classification? Introduce ImageNet and show example
  Image classification is a fundamental concept in deep learning. It identifies objects in an image by training a model through experimentation with labeled images.
  This chapter will teach you the process of identifying objects in an image and introduce you to ImageNet, along with several examples.
• Classification using a single linear threshold (perceptron)
  Here, you will learn the process of implementing classification techniques using a single linear threshold (perceptron).
• Hierarchical representations
  Here, you will learn the process of representing deep learning models in a hierarchical structure.
• Fitting parameters using back-propagation
  In this chapter, you will learn how to find coefficients (parameters) for one or numerous models for fitting data.
• Non-convex functions
  This chapter will familiarize you with non-convex optimization functions in deep learning.
• How interpretable are its features?
  Here, you will understand how the features are interpretable.
• Transfer learning
  Transfer learning is a widely implemented deep learning approach. It is a model developed for an application that can be reused as the initial point for a model on a second application.
• Other applications I: Speech recognition
  Speech recognition is a technique to transform human speech into written text by recognizing the voice of an individual.
• Other applications II: Image Classification 
  Image classification is a computer vision task that involves training algorithms to categorize images into predefined classes.

1 Case Study

Recommendation System algorithms, simply put, suggest relevant items to users - explaining the trends of their usage across a range of industries and their central role in revenue generation.

• What does a recommendation system do?
  As the name indicates, recommendation systems assist you in predicting the future preference of any product and recommending the best-suited items to users.
  In this chapter, you will understand the procedure to utilize a recommendation system to choose the best products for users.
• So what is the recommendation prediction problem? And what data do we have?
  The technique where the system predicts whether an individual or a business likes the product (a classification problem) or the reviews or ratings by them (a regression problem) is known as the recommendation prediction problem.
• Using population averages
  Here, you will understand the procedure for using population averages.
• Using population comparisons and ranking
  Here, you will understand the procedure for using population comparisons and ranking.

1 Case Study

Collaborative filtering is an aspect of recommendation systems with which we interact quite frequently. Upon collecting data on the preferences of multiple users, collaborative filtering makes predictions for the choice of a particular user.

• Personalization using collaborative filtering using similar users
  Here, you will understand the procedure to use collaborative filtering with the help of similar users.
• Personalization using collaborative filtering using similar items
  Here, you will understand the procedure to use collaborative filtering with the help of similar items.
• Personalization using collaborative filtering using similar users and items
  Here, you will understand the procedure to use collaborative filtering with the help of similar users and items.

1 Case Study

As suggested by the name itself, personalized recommendations work to filter out recommendations that are personally relevant for a user, based on their browsing trends, etc.

• Personalization using comparisons, rankings, and user items
  Here, you will learn how to utilize personalization recommendations with the help of comparisons, rankings, and user items.
• Hidden Markov Model / Neural Nets, Bipartite graph, and graphical model
  The Hidden Markov Model (HMM) is a statistical Markov model in which the system being modeled is regarded as a Markov process with hidden/unobserved states.
• Using side information
  This chapter will familiarize you with the procedure to use side information with the assistance of Meta-Prod2Vec.
• Building a system: Algorithmic and system challenges
  This chapter will familiarize you with the procedure to make a system considering algorithmic and system challenges.

In this module of the MIT Data Science course, you will get to know what networks are and how we can represent networks with their practical use-cases around us.

• Introduction to networks
  You can define a network as a group of two or more computer systems linked together using several hardware components, such as hubs, switches, and more.
• Examples of networks
  In this chapter, you will gain an understanding of all the examples of networks.
• Representation of networks
  This chapter will familiarize you with the procedure to represent networks.

1 Case Study

In the next module of the MIT Data Science online course, you will learn about the standard descriptive measures of a network, such as centrality, closeness & betweenness, and standard stochastic models for networks, like Erdos-Renyi, preferential attachment, infection models, notions of influence, etc.

• Centrality measures: degree, eigenvector, and page rank
  This chapter will familiarize you with the procedure to implement centrality measures, such as degree, eigenvector, and page rank.
• Closeness and betweenness centrality
  Here, you will gain an understanding of closeness and betweenness centrality.
• Degree distribution, clustering, and small world
  Here, you will gain an understanding of Degree distribution, clustering, and the small world.
• Network models: Erdos-Renyi, configuration model, preferential attachment
  The Erdos-Renyi model assists you in creating random networks or graphs on social networking. The configuration model is a technique to generate random networks from a given degree sequence. Preferential attachment is a method in which new network members attempt to establish a connection with the more prevalent existing members.
• Stochastic models on networks for the spread of viruses or ideas
  Here, you will gain an understanding of stochastic models on networks for the spread of viruses or ideas.
• Influence maximization
  The problem of identifying a small subset of nodes (seed nodes) in a social network that may maximize the spread of influence is called influence maximization.

1 Case Study

Here, you will get to know how to use graphical models to estimate and display a network of interactions.

• Undirected graphical models
  In this chapter, you will learn about undirected graphical models.
• Ising and Gaussian models
  Ising model specifies the joint probability distribution of a vector to understand phase transitions. A Gaussian model is a two-dimensional normal distribution of the concentration in the crosswind and vertical directions centered around the downwind axis from the initial point.
• Learning graphical models from data
  Here, you will gain an understanding of several graphical models from data.
• Directed graphical models
  A directed graphical model refers to the probability of random variables into a product of conditional probabilities, available for every node in the graph.
• V-structures, “explaining away,” and learning directed graphical models
  Here, you will understand more about directed graphical models, V-structures, and “explaining away”.
• Inference in graphical models: Marginals and message passing
  This chapter will teach you about inference in graphical models, such as Marginals and message passing.
• Hidden Markov Model (HMM)
  This chapter will brush your previous knowledge of the Hidden Markov Model (HMM).
• Kalman filter
  The Kalman filter algorithm is used to provide estimates of some unknown variables, given the measurements are observed over a particular period.

The module covers :

• Overview of ChatGPT and OpenAI
• Timeline of NLP and Generative AI
• Frameworks for understanding ChatGPT and Generative AI
• Implications for work, business and education
• Output modalities and limitations
• Business roles to leverage ChatGPT
• Prompt engineering for fine-tuning outputs
• Practical demonstration and bonus section on RLHF

The module covers :

• Mathematical Fundamentals for Generative AI
• VAEs: First Generative Neural Networks
• GANs: Photorealistic Image Generation
• Conditional GANs and Stable Diffusion: Control & Improvement in Image Generation
• Transformer Models: Generative AI for Natural Language
• ChatGPT: Conversational Generative AI
• Hands-on ChatGPT Prototype Creation
• Next Steps for Further Learning and understanding