REVIEWS 3 and 4

The questions that follow are based on the lecture topics listed below

1. Week 3: Domination
2. Week 4: Statistics

Domination

• What is a Pareto frontier?
• Give a framework for a general evolutionary algorithm
• List some of the domination methods used in multi-objective optimization
• How does binary domination work?
• How does indicator based domination work?
• In the following example, we are attempting to (a) minimize Objective-1, (b) minimize Objective-2, and (c) Maximize Objective-3. For this case, would binary domination work?
  • If yes, list the non-dominated rows.
  • If no, explain why.

        Row ID  Objective-1     Objective-2     Objective-3
        1       10             15           20
        2       12             20           22
        3        8             18           20

• How is the dom score computed in indicator based domination?
  • True/False: Lower the dom score, the better.
• In the following example, we are attempting to (a) minimize Objective-1, (b) minimize Objective-2, and (c) Maximize Objective-3.
  • Compute the dom score.
  • Sort rows based on the dom score

        %cylinders   <weight   >acceltn   >mpg
        8            4425      11         11
        8            4955      11.5       10
        8            4464      11.5       10
        8            4464      12.5       10
        4            1985      21.5       40
        4            2085      21.7       40
        4            2130      24.6       40

Evaluation measures for predictors

1. Define (with corresponding formulae) the following measures
2. RE, MRE, MMRE, medMRE, PRED(30).

Statistics

1. Define and distinguish the following:
   • Effect size and Significance tests
   • Parametric, Non-Parametric tests
2. Cohen's delta is a non-parametric effect size test. Explain.
3. Explain this statement in terms of gaussian distribution: "Any significance test reflects on the overlap between distributions"
4. For question 3 (above), draw 3 diagrams with pairs of distributions depicting the differences
5. Explain how the Scott-Knott test works.
6. In the innermost loop of scott-knott test, what kind of hypothesis would you apply?
7. What are the similarities between scott-knott tests and unsupervised discretization
8. What are the differences between scott-knott tests and unsupervised discretization

K-NN

1. KNN is called a lazy learner (and decision tree learners are not lazy). Why?
2. KNN is called an instance-based learner (while decision trees are model-based). Why?
3. Explain the following statements
4. KNN can be simply extended to steam mining while decision trees, not so much.
   • KNN is faster than decision trees, for training, but very much slower when tested on new data.
   • When k=1, the kernel function is irrelevant.
   • KNN can be optimized with some stochastic sampling
   • Any stochastic sampling method has to be certified with experimentation
5. What does the kernel function do in KNN?
6. What are linear, median, triangular kernels? Explain with a numeric example.