Machine Learning in Practice

This course is built around some practical cases; datasets are included.

For each case, we go through the entire life cycle of a machine learning project:

• Exploring, cleaning, and preparing data;
• Selecting a learning method to match the task (linear regression for regression, random forest for classification, K-average and DBSCAN for clustering);
• Learning with the use of the selected method;
• Outcome assessment;
• Model optimization;

A part of the course will be devoted to discussing practical tasks that trainees deal with, which can be solved by using reviewed methods.

• Understand what tasks can be solved with the help of machine learning (and find out that Big Data is just a subsection, not a mandatory requirement).
• Learn how to utilize basic methods of machine learning and how to use fast prototyping tools to answer the question, “Can you evaluate an actual return from possible implementation?”
• Highlight data that should be collected and what can be required from it in the near future. Why “we want to store petabytes” – it’s not always just a whim.
• Get prepared for more complex themes, particularly to complete solutions to real complex business problems.
• See how exactly machine learning fits with classical analytics. In particular, make sure that it’s unnecessary (or even harmful) to dismiss all existing analysts for concept implementation.

1. Task overview (theory - 1 hour)
   Tasks that are better solved by machine learning. What will happen if instead of a Data Scientist you hire a non-specialist in a given domain (just a developer/analyst/manager), expecting that they will learn everything in the process.
2. Preparing, cleaning, and exploring data (theory - 1 hour; practice – 1 hour)
   How to gain insight into data provided by business (and find whatever order in it at all). Processing steps. What can and should be done by domain analysts, and what should better be done by a Data Scientist. Priorities in solving a specific task.
3. Classifiers and Regressors (theory - 2 hours; practice – 2 hours)
   Practice – well formalized tasks with prepared data. Differences between tasks (binary/nonbinary/probabilistic classification, regression), redistribution of tasks across classes. Examples of practical tasks classification.
4. Clustering (theory - 1 hour; practice – 2 hours)
   Where and how to do clustering: exploring data, task setting check, and validation of results. Which cases can be reduced to clustering.
5. Model evaluation (theory - 1 hour; practice – 1 hour)
   Business metrics and technical metrics. Metrics for tasks of classification and regression, error matrix. Internal and external metrics of clustering quality. Cross validation. Overfitting.
6. Optimization (theory - 2 hours; practice – 2 hours)
   What makes one model better than another: parameters, traits, and ensembles. Parameter management. Traits selection practice. Overview of tools for searching best parameters/traits/methods.
7. Neural networks (theory 2 hours; practice – 2 hours). Gradient descent, backpropagation, generative neural networks, convolutions, recurrency, batch normalization, dropout, activation functions, batching
8. Graphs, reports, dealing with real-life tasks (theory - 1 hour; practice – 2 hours).
   How to visualize and present results. Semi-automated tests, process control points. From real-life tasks to complete R&D process (“R&D in practice”) – reviewing and analyzing tasks from the audience.
9. Data science interview questions and answers (theory – 1 hour)

Total: theory 12h, practice 12h