Visualize ICLR 2023 OpenReview Data

ICLR 2023 Paper submission analysis from https://openreview.net/group?id=ICLR.cc/2023/Conference

Requirements

• Install requirements

pip install wordcloud nltk pandas imageio selenium tqdm

• Download several nltk-related packages for language processing

import nltk
nltk.download('punkt')
nltk.download('averaged_perceptron_tagger')
nltk.download('wordnet')
nltk.download('stopwords')

• If you got anything wrong when calling webdriver.Edge('msedgedriver.exe'), you can

  • Delete msedgedriver.exe since it may only work on my computer (Windows 10)

  • Install Microsoft Edge (Chromium): Ensure you have installed Microsoft Edge (Chromium). To confirm that you have Microsoft Edge (Chromium) installed, go to edge://settings/help in the browser, and verify the version number is Version 75 or later.

  • Download Microsoft Edge Driver: Go to edge://settings/help to get the version of Edge.

  • Navigate to the Microsoft Edge Driver downloads page and download the driver that matches the Edge version number.

  From https://stackoverflow.com/questions/63529124/how-to-open-up-microsoft-edge-using-selenium-and-python

Crawl Data

1. Run crawl_paperlist.py to crawl the list of papers (~0.5h).

python crawl_paperlist.py --year 2023

Paper List

crawl_paperlist.py will miss several papers for some errors. The full paper list are as follows:

• Year 2022 (3,407 submissions in total):

  • sources/ICLR-2022.csv

  • sources/ICLR-2022.md

• Year 2023 (4,966 submissions in total):

  • sources/ICLR-2023.csv

  • sources/ICLR-2023.md

Visualization

Keywords and Title

• Keywords Frequency The top 50 common keywords (uncased) and their frequency:

• Top-10 Ranking between 2022 and 2023 (full list please refer to keywords.md)

Keyword	2022	2023
reinforcement learning	1	1
deep learning	2	2
representation learning	4	3
graph neural network	3	4
transformer	5	5
federate learning	7	6
self-supervised learning	6	7
contrastive learning	10	8
robustness	9	9
generative model	8	10

• Ranking Changes of Top-50 keywords

Keywords	2022	2023	rank $\uparrow$
large language model	208	32	176
diffusion model	173	14	159
offline reinforcement learning	59	20	39
sparsity	85	49	36
adversarial training	19	47	-28
differential privacy	45	23	22
fairness	43	22	21
model compression	61	41	20
domain generalization	55	36	19
time series	58	40	18

• Keywords Cloud The word clouds formed by keywords of submissions show the hot topics including deep learning, reinforcement learning, representation learning, graph neural network, etc.

• Title Keywords Frequency The top 50 common title keywords (uncased) and their frequency:

• Top-10 Ranking between 2022 and 2023 (full list please refer to title.md)

Title	2022	2023
representation	1	1
graph	3	2
data	6	3
reinforcement	2	4
transformer	7	5
training	5	6
image	10	7
efficient	9	8
language	15	9
federate	14	10

• Ranking Changes of Top-50 Title keywords

Title	2022	2023	rank $\uparrow$
mask	325	45	280
diffusion	132	25	107
base	76	36	40
visual	61	38	23
offline	55	34	21
attack	25	46	-21
vision	64	44	20
generation	36	17	19
adaptive	45	32	13
knowledge	38	26	12

• Title Keywords Cloud The word clouds formed by keywords of submission titles:

Related projects

• https://github.com/evanzd/ICLR2021-OpenReviewData
• https://github.com/fedebotu/ICLR2022-OpenReviewData
• https://github.com/EdisonLeeeee/ICLR2022-OpenReviewData