cscorley / dissertation Goto Github PK

For random seed, show before/after of two different seeds

• Introduction
• Terminology
• Motivation
• Research Goals and Problems
• Methodology
  • Why changesets?
  • Datasets and Benchmarks
  • On LDA: random seeds, batch, and online?
• Organization

• Intro
• Motivation
• Background
• Study design
  • Approach
  • Evaluation
• Results
  • RQ1
  • RQ2
• Discussion
  • RQ1
  • RQ2
  • Situation
  • Files, classes, & methods
• Threats

• Motivation
• Background
• Study design
  • Approach
  • Evaluation
• Results
  • RQ1
  • RQ2
• Discussion
  • RQ1
  • RQ2
• Threats

• Discussion
• Future work
• Last words

Historical should have more success than either of the others

add highlights to the config result tables for configs used in other areas of the work

currently only set once, but should reset every LDA init

• Write a "rolling synopsis" to help organize and express the current state of things (Authoring pg 53)
• send it to Nick ASAP
• consider sending to Carver?

Possibly split up the combination chapter (after reorg) into two separate ones. It seems there's plenty of need to separate model optimization and corpus optimization given how much data I have. Blocked by #4

Much of the table/figure building is spread across multiple notebooks and it's becoming difficult to keep everything in order. These need to be refactored so everything is systematic and defensible

RQ look like dogcrap

Would like one boxplot to remain as they are, and another boxplot that excludes outliers. I hope the latter 'zooms' on the boxplot a bit so highlight performance differences for non-outlier ranks

Many of these are likely out of date w.r.t. their order.

Should be possible to grab the current chapter & section numbers and inject them instead!

• Intro
• Motivation
• Background
• Study design
  • Approach
  • Evaluation
• Results
  • RQ1
  • RQ2
• Discussion
  • RQ1
  • RQ2
  • Situation
• Threats

yeah, i really dislike the 'snapshot' approach for DIT i used
i should have done something else

from an experimental perspective
depends both on the input data and the heuristic
where input data varies by snapshot and changesets, and heuristic varies by change counting and indexing dev documents
having only the input data differ would have been better, and its not unobtainable

(need to edit this later to clarify)

There is a need to show why a good seed or a bad seed can change outcome.

Visual idea: compare two LDA state matrices after update with each document/chunk

Right now the topics are spread across chapters. Instead, flip chapters and sections.

search for articles using abstracts

Turns out there's a lot of repeated text & ideas that would be better off served if the chapters were inverted with sections.

• Intro
• Motivation
• Background
• Study design
  • Approach
  • Evaluation
• Results
  • RQ1
  • RQ2?
• Discussion
  • RQ1
  • RQ2?
• Threats

• Intro
• Motivation
• Background
• Study design
  • Approach
  • Evaluation
• Results
  • RQ1
  • RQ2?
• Discussion
  • RQ1
  • RQ2?
• Threats