Skip to content

Latest commit

 

History

History
102 lines (66 loc) · 3.76 KB

review4.md

File metadata and controls

102 lines (66 loc) · 3.76 KB
Raw

 

home | copyright ©2019, tjmenzie@ncsu.edu

syllabus | src | submit | chat

Review 4

Ensembles

Bagging

  • What is bagging
  • Why can 10 "bags" do better than one?
  • What is the connection of bagging to cross-validation?

Boosting

  • What is boosting?
  • Why can 10 "boosts" for better than one?
  • What is the difference between bagging and boosting
  • Explain: bagging is inherently parallel while boosting is inherently sequential

Statistics (more)

T-tests

T-tests report the overlap of two normal bell shaped curves. Describe how they might be used in a 10-way cross-val experiments to rank different learners (note: your description should be high-level: no formulas required).

Using the terms standard deviation and mean, describe two such curves that t-tests would find

  • easy to distinguish. Draw those two curves.
  • hard to distinguish due to some value of the means. Draw those two curves.
  • hard to distinguish due to some value of the standard deviations. Draw those two curves.

A t-test is a parametric statistical significance test

  • what parametric assumptions re made by the t-test?
  • for what kind(s) of curves does the t-test not hold?

Bootstrap

A bootstrap is a non-parametric test of statistically significant difference:

  • What is the role of the "test statistic" in the bootstrap?
  • What is the role of "sampling with replacement" in the bootstrap?
    • Given the list of numbers [10,20,30,40] write down 5 such samples
  • Using the terms "test statistic" and "sample with replacement", describe the bootstrap (note: your description should be high-level: no formulas required).

Cliff's Delta

Cliff's Delta is a non-parametric effect size test:

  • What is the difference between an effect size test and a significance test?
  • How does Cliff's Delta sample two lists to see if one is more than trivially different to the other? Use pseudocode
  • Explain: non parametric methods are slower than parametric method since they have to using sampling, not summarize , of the data

Clustering:

Distance

  • Describe the distance measure presented in lectures and how it can do distances between vectors containing symbolic or continuous values
  • Why does that distance metric normalize numbers 0..1 min..max
  • How does that metric handle missing numerics?
  • How does that metric handle missing symbols?

Mini-batch k-means (MBKM)

  • Explain. MBKM works in batches to move centroids less and less.
  • Explain: MBKM is a better choice than K-means for very large data sets

KD-trees

  • Describe the KD-tree algorithm. Assume ranking via max standard deviation and splits at mean.
  • Explain: KD-trees suffer from the curse of dimensionality

Random projections

  • Given the distance measure described above
    • describe a random projections clustering algorithm that divides 900 examples into clusters of around size 30
    • Describe how to reduce the curse of dimensionality in KD-trees
  • Given a matrix generation algorithm that files columns with gaussians pulled from (mean,sd)=0,1 describe how to reduce
    • a matrix A with m × n rows and columns
    • to a narrower matrix B with m × p rows and columns (p < n)

Aside:

For this kind of RP, it has shown that the Gaussian distribution can be replaced by a much simpler distribution such as

R[i,j] =

  • sqrt(3) with probability 1/6,
  • zero with probability 2/3,
  • and -sqrt(3) with probability 1/6.