Meeting Notes Oct

[lujiaying]

1 contains the kick-off until Oct 6 meeting noes. Let's use this new one for a cleaner meeting note trakcer.

Long-term Goals:

• Add Magic, Diablo, or others if needed
• mm-gnn design, other baselines.

[lujiaying]

Oct 7 - Oct 13

• Pokemon, Hearthstone and LoL-Skin
  • [x] @lujiaying Pokemon type2
  • I decide NOT regroup anything, as making Pokemon a highly-imbalanced dataset
  • [x] @qyccc3 Hearthstone race: remove test < 5 categories
  • [x] @qyccc3 Lol-Skin: merge test==1 categories into Others
  • [x] @qyccc3 could you also add url/website for collecting Lol-Skin ?
  • each dataset contains info.json instead of info.txt for the sake of machine-readable configure.
  • [ ] add elaboration of dataset creation process in #3 . @qyccc3 please see the comments there. There is something missing.
• Benchmarking:
  • [ ] baselines other than AG-tabular(B), AG-tabulr-mm(B), and AutoMM (details please refer to #5)
  • [x] @lujiaying scripts ready to run by Saturday
  • [x] @qyccc3 Hearthstone minion, spell on Hopper, with 8 hours, 16 CPUs, 1GPU constraint. It would be great to check with @lujiaying for first-time running.
  • [ ] please upload result.csv under Onedrive-ExpResults. For now, just upload each individual csv, and use a proper name.

Next week plans

• add automm baseline, mimic autogluon exec.py
• collect generated results into one file, upload them to overleaf. (script generate table from csv to latex)
• run pokemon, LOL
• learn dgl tutorial, learn basic concept of GCN classification. Try to construct a graph by category features.

[lujiaying]

Oct 14 - Oct 20

• [x] @lujiaying look into logs, then figure out what to do: enforce sequential training c53128597
• [x] @qyccc3 add automm baseline and exec.py , mimic autogluon exec.py. I would suggest creating a pull request for debug purposes. (ref1, ref2)
  • No worries, @lujiaying has did it
• [ ] @qyccc3 maybe need to re-run. Please also finish pokemon and Lol-skin.
• [ ] @qyccc3 we want to have result.csv uploaded into Onedrive. results numbers would be expected to upload to overleaf.
• [ ] (low-priority) @qyccc3 learn gcn-tutorial, rgcn tutorial;
  • [ ] step 1: read and analyse TabGNN, Multiplex Graph Neural Networks, www'22

[lujiaying]

Oct 21 - Oct 27

• [x] add table (accuracy, balanced ACC)
• [ ] investigate why mm is worse..: autoMM not converge?
• [x] Dataset
  • [x] datasets to delete: hearthstone all_rarity, all_cost, minions_health, minions_attack
  • [x] hearthstone rarity, we need to convert None_rarity to free
  • [x] hearthstone-Minions_race, we expect health, attack and cost as raw features, instead of processed features. (i.e. no need to have health=11~20)
  • [x] hearthstone-Spells_spellSchool, please remove health column
• [x] New Dataset: CSGo-Skins (concern: maybe too easy)
  • [x] possible label cols: skin quality, skin category
  • [x] column generation: skin price -> skin price min, skin price max