search

chengsoonong / mclass-sky

Multiclass methods for astronomical data
BSD 3-Clause "New" or "Revised" License
9 stars 4 forks source link

Louche and Ralaivola paper #121

Closed davidjwu closed 8 years ago

davidjwu commented 8 years ago

Make notes on the Louche and Ralaivola paper (pdf) and duplicate their experiments on a subset of the SDSS data set.

davidjwu commented 8 years ago

@chengsoonong The paper compares the performance of two active learning algorithms, Active-SVM and Active-BPM, with the difference being that the former is based on the Chebyshev center and the latter is based on the center of gravity.

It seems Active-SVM was implemented using the package LIBSVM (site), which has a Python interface. So I may attempt to replicate the results in the paper attained using LIBSVM. Alternatively, it appears computing the Chebyshev center reduces to solving a linear program, so attempting to do this directly using NumPy, SciPy or perhaps CVXOPT (site) may be more straightforward.

Unfortunately, Active-BPM seems to have been implemented in MATLAB, which is problematic. I've been unable to find a Python package that I can use to computes the center of gravity. From what I can gather, computing the center of gravity is not straightforward at all.

chengsoonong commented 8 years ago

You do not need to compare with Active-BPM.

I suggest comparing three algorithms in your work:

  1. Algorithm 4 in the paper, with a uniform random point as center(C^t) (issue #123)
  2. ACCPM with Algorithm 4 in the paper (issue #122)
  3. (optional) Chebyshev center using cvxpy (raise issue if you get here).

I suggest solving them in the suggested order above. Please confirm that you are happy with this plan @davidjwu, and we can close this issue.

davidjwu commented 8 years ago

@chengsoonong Yes, this plan seems reasonable.