Course: Machine Learning

(Alternate Name for course) Yes, there are other learning methods other than neural nets (and genetic algorithms).
Edward (Ned) S. Blurock
(Home Page)
E-Mail:
Edward.Blurock@forbrf.lth.se
Snail-Mail

Combustion Physics
Lund University
P.O. Box 118
SE-221 00 LUND
Sweden
Phone: +46 46 222 1402
FAX: +46 46 0885


This is all the course materials and slides (to date) for the winter semester 94 lecture on Machine Learning. The purpose of the course is to introduce a class of learning methods having more to do with (what I will call) predicate descriptions of the training examples. Neural nets are not covered because this is another course in this series and genetic algorithms are not covered (extensively) because discussion of this should be under optimization and search techniques. ID3 is used as the introduction to learning methods in general and specific problems effecting this and other learning methods (i.e. missing data, pruning) will be discussed within the framework of this method. Other methods will also be introduced: Mikalskis's AQ and Conceptual Clustering and the class of incremental concept formation algorithms, EPAM, UNIMEM and COBWEB (and maybe CLASSIT). Apart from the theoretical introduction to machine learning, experiments in parameter specification and analysis (using the ANALYSIS system) will be performed. The course is itself an experiment in "Learn by Doing".

Lecture 1: Introduction to Machine Learning

A very general overview about what one is trying to accomplish with machine learning. The emphasis of this course will be methods in which qualitative information can be extracted.
Handouts:
Lecture Notes I:
Scope of Lecture and What does one what to accomplish with machine learning
Lecture Notes II
An Example using the machine learning program Analysis
Exercises:
None
Overheads:
Intro

Lecture 2: What is needed for machine learning (The Analysis Program)

A tour of the parts of the Analysis program and at the same time an introduction of the essential parts of a machine learning calculation to produce a decision tree
Handouts:
Overview of Analysis
Exercises:
1 and 2
Overheads:
Analysis 4.0

Lecture 3: Details of the ID3 Calculation

The ID3 method is explained in more detail. First, an intuitive introduction and then a more in depth explanation based on the paper of Quinlan.
Handouts:
ID3
J. R. Quinlan: Induction of Decision Trees, Machine Learning, vol 1, p 81-106 (1986)
The Analysis Directory
Exercises:
3 and 4
The ID3 Selection measure and experiments in decision tree size.
Overheads:
Details of the ID3 Method

Lecture 4: Issues in Decision Tree Making I

Two aspects of decision tree making are explored: Missing Values and Selection Criteria. The basis of the lecture are several papers
Handouts:
J. R. Quinlan: Unknown Attribute Values in Induction, Proc. 6th Int'l Workshop Machine Learning, 1989
A review and comparison of Several Ways to deal with missing values
John Mingers: An Empirical Comparison of Selection Measures for Decision Tree Induction, Machine Learning, vol 3, p 319-342 (1989)
A review of several selection measures
Wray Buntine and Tim Niblett: A Further Comparison of Splitting Rules for Decision Tree Induction, Machine Learning (1993)
This paper was included not only because of the more exact comparisons of splitting rules, but also because of the outline of data sets and their properties.
Exercises:
Experiments with missing data (5-6)
An simplified experiment is set up exploring the the effects of substituting various values for missing attributes
Representing a polynomial for machine learning analysis (7)
This is the first step of trying to represent a polynomial system in terms of a set of parameters. The purpose is to illustrate the problems involved in representation.
Selection Measures
Calculation of Selection Measured for a few parameters (in the voting data set)
Overheads:
Missing Data
Minger Paper
Buntine-Niblett Paper

Lecture 5: Issues in Decision Tree Making II

The Liu and White paper is used to explain some of the problems in attribute selection and why the quality of selection measures is even an issue to be discussed. Another issue is that of pruning the decision tree. The paper of Mingers is used because several pruning methods are introduced
Handouts:
W.Z. Liu and A.P. White: The importance of Attribute Selection Measures in Decision Tree Induction
An example of a study with random attributes and random selection criteria
John Mingers: An Empirical Comparison of Pruning Methods for Decision Tree Induction
Five methods of pruning are discussed
Exercises:
Polynomial Analysis w.r.t. Calculation Times (9-10)
Overheads:
Liu and White Paper
Pruning

Lecture 7: Bayesian Statistics and Decision Theory

Since one of the purposes of machine learning is to make decisions, bayesian statistics are introduced with examples of the expectations of the accuracies of predictions. Part of the class is used to discuss the polynomial paramters that were created.
Handouts and References:
"Chapter 19: Bayesian Inference", Introductory Statistics, Wonnacott, Thomas H. and Wonnacott, Ronald J.
Prior and Posterior probabilities and Likelihood functions are introduced with examples first from binary decisions and then using normal and binomial distributions.
Overheads:
Bayesian Inference
Exercises and Discussion:
Continuing work on the Polynomial problem: Discussion of what set of descriptors one could make to describe the special set of 30 polynomials . Assignments were given to create these descriptors.

Lecture 8: Inductive Learning: Generalization and Specialization

The concepts and notation of inductive learning in terms of generalization and specialization are introduced. This is used to put the inductive learning concepts on a bit more formal level and to introduce the star concept and the AQ Algorithm.
Handouts and References:
A Theory and Methodology of Inductive Learning, Michalski, Ryszard S. in Machine Learning: An Artificial Intelligence Approach

Overheads:
Generalization and Specialization
Star And AQ
Exercises and Discussion:
Continuing work on the Polynomial problem: The data sets have been submitted and a short intuitive analysis (using Analysis) is looked at.
The data sets and experiments:

Lecture 9: Inductive Learning: Generalization and Specialization

Half of the lecture is devoted to explaining the star technique in more detail and (on the blackboard) giving examples of the implementation in Analysis (i.e. the descriptions as a set of predicates). The other half of the lecture is a more intensive discussion of the polynomial problem.
Handouts and References:
Learning from Observation: Conceptual Clustering, Michalski, Ryszard S. in Machine Learning: An Artificial Intelligence Approach

Overheads:
Conceptual Clustering
Exercises and Discussion:
Continuing work on the Polynomial problem: A more intensive discussion of the polynomial problem

Edward.Blurock@forbrf.lth.se