Midterm: (Time: 8:30 - 9:50, February 24th (Tuesday); Location: Building 200/Room 105)

It's a closed book exam. However, you are allowed to bring one piece of letter sized and double sided notes to the exam.

The midterm is based on:

• up to Probability Based Classification.
• Labs 1 to 5
• Assignments 1 to 3
• Week 1 to Week 6 (inclusive)

Topics included:

• Introduction
• Data
  • Why is it important to understand the data collection?
  • Attribute types (and later, how they affect the learning algorithms)
  • Common issues and how they are handled: missing values, noisy data
  • Data aggregation, reduction and transformation
• Information Based Learning
  • product: decision tree
  • how to build it: recursive algorithm, greedy algorithm
  • information in "information based": entropy, GINI, mis-classification rate
  • how to use it
  • issues (and how to handle them)
    • bad split situation (use information gain ratio instead of information gain)
    • overfitting (pre-pruning, post-pruning)
    • continuous descriptive attribute (discretization - equal width, equal depth, information based)
    • continuous target attribute (regression tree - using variance instead of entropy)
  • model ensembles (decision forest instead of tree)
• Similarity Based Learning
  • feature space (data format)
  • similarity (distance) metrics
  • how to use it: nearest neighbor algorithm
  • issues (and how to handle them)
    • noisy data (K nearest neighbor algorithm)
    • feature difference (normalization)
    • continuous target attribute (average instead of class label)
    • too many descriptive attributes (dimension reduction)
  • improve efficiency - indexing (such as K-D tree), pre-screening
• Probability Based Learning
  • some statistics concepts
  • how to get the probabilities
  • Bayes' Theorem
  • how to use probability based classifier in general?
  • Naive Bayes' classifier, and its BIG assumption
  • Bayesian Belief Networks
    • what does it look like?
    • Markov blanket
    • how to use it?
    • how to build it?
  • issues (and how to handle them)
    • insufficient data (smoothing)
    • continuous descriptive attribute (probability density function instead of a fixed set of probabilities)
    • missing parent attribute values - hidden variables (consider all possible values and their probabilities)

Past exams: