DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
UNIVERSITY OF CALIFORNIA, SAN DIEGO
CSE 291: Data mining and predictive
analytics
Spring 2012
Please ask
questions on Piazza.
Use the same link to sign up.
Lecture
notes are in the file dm.pdf, which is
updated regularly. Each week's assignment is at the end of the
relevant chapter.
CSE 291 in
the spring quarter of 2012 will be a graduate-level lecture course
devoted to current methods for data mining and predictive
analytics. (The course will be renumbered CSE 255 in the future,
but has not been renumbered yet.) No previous background in
machine learning is necessary, but all participants should be
comfortable writing programs and following arguments that use
basic calculus and linear algebra. Although the course will use
some mathematics, the focus will be on useful skills more than on
theoretical foundations. The course meets once a week in the evening for ten weeks.
Meetings will be on Tuesdays from 6:30pm to 9pm in room 2154 in
the CSE building, beginning on Tuesday April 3, 2012. The last
lecture will be on Tuesday June 5. The final exam will be on
Tuesday June 12.
date
|
topics
|
lecture notes
|
April 3
|
Course outline, limitations
of data mining, linear regression, overfitting,
regularization
|
Chapters 1 and 2
|
April 10
|
Linear and nonlinear support
vector machines (SVMs), cross-validation
|
Chapters 4 and 5
|
April 17
|
Confusion matrices, ROC
curves, classifier learning when one class is rare
|
6
|
April 24
|
Predicting probabilities,
logistic regression
|
7
|
May 1
|
Confidence intervals, cost
matrices, making rational decisions
|
8
|
May 8
|
Training classifiers for
documents
|
11
|
May 15
|
|
|
May 22
|
|
|
May 29
|
|
|
June 5
|
|
|
June 12
|
Final exam in CSE room 2154, 6:30pm to 9pm
|
|
For the 2011
version of the course, see http://cseweb.ucsd.edu/users/elkan/291spring2011. Topics for 2012
will include most of the following:
- Linear
regression, regularization, selection and creation of predictors
- Cleaning and
recoding data
- Classifier
learning: support vector machines and random forests
- Learning when
one class is rare, sensitivity/specificity tradeoffs
- Practical
issues, including overfitting, leakage, and cross-validation
- Making
optimal decisions based on probabilities and costs
- Maximizing
lifetime value and reinforcement learning from historical data
- Recommender
systems and collaborative filtering
- Reconciling
privacy with data mining
- Mining
collections of text documents
- Analyzing
social networks and link prediction
- A/B
testing and systematic experimentation.
Enrollment is unrestricted. Everyone is
welcome. Community members will register through UCSD Extension concurrent
enrollment. Participants may take the course
for a letter grade (preferred), or for a
satisfactory/unsatisfactory (S/U) grade. All participants should
register for exactly four units, and are expected to do the
assignments, quizzes, and final exam. There will not be any
midterm.
The instructor
is Charles Elkan
(Professor), whose office is in the CSE building, room 4134. Feel
free to send email to
arrange an appointment, or telephone (619) 379-9852. The teaching
assistant is Spencer
Bliven.
For the 2011
version of this course, see http://cseweb.ucsd.edu/users/elkan/291spring2011.
The course will not follow a textbook, and no specific book is
required. See here for some suggested
books.
Each week there
will be a hands-on assignment due in class the next week. Assignments
will include pointers to datasets. You should do each assignment
in a team of two or three. You are free to keep the same partner
for multiple assignments, or to switch. You should look for
intellectual diversity in whom you choose as a partner.
Specifically, people from the same company should not work
together. Students from outside CSE should pick partners who are
CSE students, or similar.
For each
assignment, each team should turn in a brief joint report. Each
report should be single-spaced and include figures, tables, and
citations as appropriate. Do not include appendices or listings
of code. Grades will be based entirely on the joint reports.
Students can
choose which software to use. The recommended package, called R, is completely
free and open-source. RStudio
is an easy-to-use free interactive environment that makes R comparable to
Matlab (which is recommended also, but is not free). The New York
Times had a good article about
R. The
Rapidminer package is a
100% Java interactive data mining environment. It is easier than R
for non-programmers to use, but R is more flexible and scales
better to large datasets.
Starting on
Tuesday April 10, there will be a seven minute quiz at the start
of each class meeting. The first few minutes of each class will be
for answering questions and returning previous work. The purpose
of the quiz is to encourage participants to arrive on time, and to
review the material from the previous class meeting. You will do
each quiz jointly with one partner. Talking is encouraged, and
overhearing other students is fine. After each quiz, there will be
a brief class discussion about what the correct answer is.
The lowest quiz score will be dropped. If you miss a quiz, its
score is the one that will be dropped. Quizzes and the final exam
will be open-notes. This means that you may bring your own notes,
a printed copy of the class lecture notes, and copies of your own
previous assignments and quizzes. Since no book is required for
this class, please do not bring books. Electronic devices are also
not allowed.
Most recently
updated on May 8, 2012 by Charles Elkan, elkan@cs.ucsd.edu