daffidwilde / kmodes-init-paper Goto Github PK

If making this repo public (#44) then we need to make sure the notebooks are presentable

The abstract is not appropriate for the results of the paper anymore and needs updating.

In addition, the opening paragraphs of the text should be revised to be punchier and focus more on the success of the new method.

The current title is long and reads weird.

"A novel game-theoretic initialisation process for the k-modes algorithm using the hospital-resident assignment problem"

• Start writing section on preference lists
  • Best, worst, random
  • Is there scope to incorporate prior knowledge this way?
  • Could there be a matching where we always have the initialisation from Huang's and Cao's methods?
• Get some numerics down for all of these implementations
  • Separate scripts for matching initialisations?
• Write definitions
• Light editing of algorithms and examples

This change should be concise as this entire subsection is getting gutted re: #48.

If the fitted distributions are there for their parameters, bimodal models need to be figured out in Python. Otherwise, if they are there to accentuate the shape of the histogram, is a KDE sufficient?

• DOI for top percentile data, source code and notebooks
• Write README for source code/data archive (#44)
  • Details of directory tree
  • Instructions for installing conda environment

• Genetic algorithm for suitable dataset generation; identifying datasets primed for a particular algorithm against its rival
• Is there a computational gain from using this process on large (complex) datasets? Not only on real data but large artificial datasets
• Objective graphs
  • Plot of the function itself
  • Plots over time
• Start exploring preference lists; this needs to be done in tandem with everything else
• Continue with the data analysis
  • A presentation would be good (for Kendal)

• Mention the matching library when the HR algorithm is introduced (citation)
• Create a release for my fork of kmodes and update the environment.yml
• Mention the release in the results section (DOI)

• Fine tuning preference lists to hopefully get better results
• Is there scope to introduce 'expert' knowledge? Or is this just an analogue of weighting the data?
• What were the characteristics of the data on good runs?
• Change the way results are gathered. We don't want to classify/predict anything - we want to find intrinsic properties
• Does starting at a smarter point make the final cost 'better'? Then we can talk about its quality as a predictive model
• The bug in the capacitated algorithm needs to be investigated as it has performed better for predicting

It would be great to have a diagrammatic example to go alongside this paper (or the respective chapters in the thesis) so dump it/any discussion about it here.

• Table (chart) showing notation
• Add m, N to example 1
• Blocking up algorithms? Makes them easier to compare and read themselves. An alternative could be to give a higher level version in the main text and the full version as an appendix.
• Side-by-side steps for Huang/Cao examples
• So what? Is there something this can do that the other intialisation processes can't?