Empirical Methods in Machine Learning & Data Mining
CS578
Computer Science Department
Cornell University
Fall 2003

General Information    Lecture Notes    ML Links    Assignments    Project

Announcements:

missing values and feature selection slides available here (slides.ppt)

The Final Exam is open book, open notes, but you are not allowed to use any devices that might be wireless.  That means you cannot use laptops, PDAs, etc. during the exam.

For regrades on the mid-term it's best to talk to the individual who graded that question:

Rich graded  OG &  CA
Alex - SVM & CV
Casey - ANN
Cristi - kNN
Radu - CO & DT

For the final project, class labels are in the first column, not the last column.  The class labels have all been set to "0" for the test data.

Predictions are due midnight Sat Dec 6.  We will not start assessing late penalties until Monday Dec 8.  Penalty is -5 for predictions emailed to us Mon Dec 8, -10 for predictions emailed Tue the 9th, -15 for Wed the 10th, and -20 for Thu the 11th.  No predictions will be accepted after Thu.  The report *must* be handed in by Thursday.  Late reports will not be accepted --- no exceptions!

We will ignore predictions that do not obey the submission format!  No extensions for submitting the wrong format.  (NOTE: you should test email the predictions to yourself to verify that they are being sent as attachments, and not as text inside the body of the email.  If possible, send the files gzip'd (unix) or winzip'd (windows).)

The report for the project is due Thursday, Dec 11 at noon!  Give the report to my admin Cindy Robinson in Upson 4146, or to one of the TAs.  Unlike homeworks,  the final report is graded on how clear, concise, and well organized it is.  

The project is out. Predictions due Dec 6. Text is here (project.ps) and data is here (project.data.tar.gz) and the code for perf5 is here (perf5.c).  

NOTE: when you submit predictions for accuracy for the final project, you should adjust your predictions so that when we use a 0.5 threshold it does what you wanted, or, apply the threshold yourself and just send us the test cases classified as 0 or 1.

Hw3 is due Nov 25. Text is here (hw3.ps) and data is here (hw3.data.gz)

Mid-term take-home exam due 5pm Nov 7.  (midterm.txt)

HW2 now due Fri 10/31/03 at 5pm.  Small extra credit for handing it in on Thu.  As before, we will be generous with late penalties so it is better to hand in an excellent HW a little late than to hand in a poor HW on time.  Happy Halloween!

RMSE baseline computation

Radu's Friday Office Hours have changed !

Have a good Fall Break!

Homework 2 is due in 3 weeks at start of class on Thursday 10/30/03.  This is a computationally expensive assignment so don't wait until the last minute to start.  Train the neural nets now while learning about SVMs.  Download data here (hw2.data.gz).  Download assignment here (hw2.ps).  

Measurements that might be useful for hw1. Handed out in class Tue Sep 23 (.ps)

Small update to cgwin install instructions.  New instructions here.  If it's already working for you don't bother.  

You should "gunzip cs578.hw1.tar.gz" before untaring file.

Homework 1 available.  Due in 2 weeks at start of class on Thursday 9/25/03
Instructions for installing IND on Windows machines are here

New Room Starting 9/02/03: Thurston 205

Time and Place

Personnel

Go to top

General Information

Course Description:
This implementation-oriented course presents a broad introduction to current algorithms and approaches in machine learning, knowledge discovery, and data mining and their application to real-world learning and decision-making tasks. The course also will cover empirical methods for comparing learning algorithms, for understanding and explaining their differences, for exploring the conditions under which each is most appropriate, and for figuring out how to get the best possible performance out of them on real problems.

Tentative Course Syllabus

Textbooks:
Machine Learning by Tom Mitchell
The Elements of Statistical Learning: Data Mining, Inference, and Prediction by T. Hastie, R. Tibshirani, J. Friedman.

Optional references:
 Pattern Classification 2nd edition by Richard Duda, Peter Hart, & David Stork

Grading policies:

• 20% Midterm (take home)
• 20% Final (open book in class)
• 30% Assignments (individual)
• 30% Final Project (group project comparing various learning methods on two test problems)
• Bonus points for class participation
• Homeworks, the take-home mid-term, and the final exam must be your own work. For homework, it is OK to talk with other students about the assignment, ask each other questions, and in general learn from each other. But the homework you hand in must be your own work. If other students gave you significant help with your homework you should briefly acknowledge them in what you hand in.

Academic integrity policy

Go to top

Lecture Note

1. Decision Trees (.pdf)

2. Support Vector Machines (.pdf)

3. KNN (.pdf)

4. Missing Values and Feature Selection (.ppt)

5. Bagging-Boosting (.ppt)

6. Performance Measures (.pdf)

7. Clustering and Unsupervised Learning (.pdf) 

8. Multi-Task Learning (.ppt) 

Go to top

Assignments

Download homework #1.  Due Thu 9/25/03 by start of class.
Instructions to install IND on a windows machine.

Go to top

Final Project

• You are encouraged to work on the project in groups of 1-4. Please register your group once it is formed. If you wish to work on the project with other students, but cannot find partners, please let us know and we'll try to match you up. It is OK to work on the project alone if you prefer.
• The final project is a mini competition. We'll hand out a training set and a final test set. For the training set you will know the target values so that you can do supervised learning using decision trees, k-nearest neighbor, artificial neural nets,etc.  You are allowed to use any of the learning methods we discuss in class. You can also combine several different learning methods if you wish. For the test set you will not know the target values. Your job is to train models on the training set, and use the best models you can build to make predictions for the test set. You then send us your predictions on the test set, and we'll measure your "accuracy". The projects getting the highest performance on the test set "wins". Part of the grade for the project will be based on how well your models perform on the test set. The final grade for the project also will take into account what methods you tried, how well you tackled each problem, and the quality of the write-up.  Read the project handout for more details.
• The final project is 30% of the course grade. Exceptional projects may get extra credit.
• Final report. Due TBA

Go to top

ML Links

• The UC-Irvine ML Dataset Archive | The UC-Irvine KDD archive | more datasets
• ML Projects
  • The WEKA Machine Learning Project (code for many ML algo's, as well as some datasets)
• Journals
  • Machine Learning | Journal of ML Research | Data Mining and Knowledge Discovery | Journal of AI Research | AI Magazine | IEEE Neural Networks Council (several journals are connected to this page)
• Knowledge Discovery in Databases

• Other University ML classes
  • Wisc Madison More External AI References
• Pointers to ML Courses
• Machine Learning Benchmarking
• AI Bibliography Server | Neural Networks Bibliography Server (Austrian AI Institute)
• AI Resources (Canadian NRC Server)
• Aha's ML Links
• Stuart Russell's:AI on the Web (loads of links)
• Reinforcement Learning Repository

Go to top