Nov 4 - Nov 10

Jiaying: continue running the other baseline models; prepare the outline of the paper
Yongchen: run AG-tab and AG-MM for the remaining datasets; paper writing for Section 3 Dataset Creation. Several references: https://dl-acm-org.proxy.library.emory.edu/doi/pdf/10.1145/3477495.3532019, https://arxiv.org/pdf/2202.11684.pdf (or other SIGIR'22 resource papers)
Shifan: paper writing for Section "related work". It would be great if we find out any baseline method that we may need to include. I tried to add what I know so far.

Dataset label leakage

Hearthstone-All-set: column id contains set information. -> id to anonymous_id by: uuid.uuid3(uuid.NAMESPACE_DNS, id)
- [x] re-run all-set

Baseline changes

do we even need two-modality methods?
- No. https://aclanthology.org/P18-1208.pdf. Following this work, we can show two models for each modality!!
tab-img now using SWIN+MLP fusion, maybe change to LightGBMLarge + CNN
tab model change from AG-tab into LightGBMLarge
clip: "data.categorical.convert_to_text": True, "data.numerical.convert_to_text": True. We need to change to False.
AutoMM: fuse change from MAX into concat (or other choices, because now they are too similar to each other..)

Read survey, and then discuss if we need more baselines...

Use Yongchen's result for ag-tab-mm for hearthstone-all-set Teaser picture show what it looks like for Hearthstone table 1: fill out last row.

Datast analysis

class label imbalance ratio..
tabular
- proportion of missing values, etc... (AMLB)
text
- total number of words, distinct words, (https://aclanthology.org/P18-1208.pdf)
image
- image size, file size
- RGB @qyccc3 any non-model analysis
dimension reduction to show feature-label correlation. (tabular features -- pca --> 2 dim vector).
- if PCA requires numerical input, Use our feature pipeline to generate one-hot encoded categories @lujiaying
- tabular raw feature v.s. trained tabular nn emb
- text using bag of words to pca v.s. trained CLIP emb
- image raw pixels can be applied to pca v.s. trained CLIP emb
- which dataset is more balanced?

lujiaying / MUG-Bench