Course Title: Machine Learning (Master’s Bologna Study Program)

Lecturer:

Prof. Dr. Igor Kononenko

Assistants: Dr. Petar Vračar, Dr. Matej Pičulin

Main Goal of the Course:

To understand the basic principles and methods of machine learning.

Content:

  1. Overview of machine learning methods
  2. (Machine) learning and (artificial) intelligence, early successes of machine learning
  3. Basic principles of machine learning and investigative algorithms for machine learning subproblems
  4. Advanced methods for attribute evaluation
  5. Introduction to artificial neural networks: Hopfield NN, RBF, multi-layer NN, (informative deep NN only)
  6. Unsupervised learning (informative only classification), association rules
  7. Knowledge representation (informative only ILP), Bayesian learning: Bayesian classifier, Naive Bayes and non-naive Bayes, semi-naive Bayes, Bayesian networks, TAN
  8. Probability calibration and reliability evaluation of individual predictions, combining machine learning algorithms
  9. Visualization and interpretation of individual predictions
  10. Learning from data streams, active learning
  11. Basics of learnability theory

Topics not covered in the course:

(multitask learning, semi-supervised learning, reinforcement learning, constructive induction, learning plans, equation system discovery, image mining, graph and network mining, COLT, time series analysis, spatial data mining, user profiling, recommendation systems, inductive logic programming)

Prerequisites:

  • Construction and pruning of decision and regression trees
  • Naive Bayes
  • K-nearest neighbors
  • Locally weighted regression
  • Learning evaluation (classification accuracy, confusion matrix, Brier score, information content of response, sensitivity, specificity, ROC curve, MSE, RMSE, MAE, RMAE, cross-validation)
  • Preprocessing of learning examples, visualization
  • Algorithm comparison, statistical tests, Bonferroni correction

Exercises:

In exercises, methods and techniques presented in lectures are practiced. Students also solve homework assignments (online quizzes, from which each student receives a grade DN, which must be at least 50% to receive a grade for the course).

Research Tasks and MLDM (Machine Learning and Data Mining) Workshop:

Each student (or up to two in a group):

  • Chooses a subfield of machine learning (except for computer vision learning),
  • Prepares a review of the chosen subfield, selects a specific problem/data and machine learning algorithms, conducts experiments, and analyzes the results,
  • Mid-semester, presents the idea of the research paper and work plan during a lecture,
  • Independently writes a report in the form of an article for the workshop (in English) and submits it for the workshop,
  • Defends the research task at the workshop (MLDM Workshop), which is organized in the last two weeks of the semester as part of the lecture.

The grade for the research task (provided the homework grade is positive) counts as the exercise grade (OC). A positive grade from the exercises is a prerequisite for taking the exam. The exercise grade is valid only for the current year. If the student does not pass the exam in the current year, they must redo the homework and research tasks in the next academic year.

Each student (or group, if the research task is carried out by multiple students) must write a report in the form of a scientific article - in English, 6-8 pages long. The format is specified by the instructions for writing an article for the workshop. The PDF version must be submitted by the deadline. At the workshop, each article will be presented in a short lecture (10 min + 5 min discussion).

Participation in the workshop is mandatory for all students.

Exam and Exercise Grade:

[ \text{Grade} = \frac{I + V}{2} ]

  • ( V = OC ) if ( DN \geq 50% ), otherwise ( V = 1 ) (fail)
  • The condition for taking the exam is ( V > 5 ).
  • ( I = ) grade for the written and possibly oral exam.

Exam:

The exam consists of a written and possibly oral part. For the written exam, one A4 sheet handwritten with a regular pencil (erasable) and signed with a pen with the student’s name and registration number is allowed (photocopies and prints are not allowed). This sheet must be submitted along with the written work.

Study Literature

Core:

BOOK:
Igor Kononenko, Matjaž Kukar: Machine Learning and Data Mining: Introduction to Principles and Algorithms, Horwood publ., 2007

PAPERS:

  • Reliability of single predictions in classification and regression (paper: Comparison of approaches for estimating reliability of individual regression predictions)
  • Calibration of probabilities (paper: Predicting Good Probabilities With Supervised Learning)
  • Explaining classification of a single instance (paper: Explaining instance classifications with interactions of subsets of feature values)
  • Learning from imbalanced data sets (Book chapter: Use of Prediction Reliability Estimates on Imbalanced Datasets)
  • Active learning (Paper: Active Learning Literature Survey)

Additional:

Ian H. Witten, Eibe Frank: Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. The Morgan Kaufmann, 1999.