cop26-27-tweets

cop26-27-supplement

Overview

This dataset contains anonymised metadata for 12,901,991 tweets collected during the COP26 (October–November 2021) and COP27 (November 2022) climate conferences. The data supports a study that combines machine learning classification, social network analysis, and qualitative close reading to analyse how climate misinformation circulates across user communities on Twitter.

No tweet text is included. User identities are replaced with truncated SHA-256 hashes. Tweet IDs are preserved for potential rehydration, though changes in the platform's API access policies mean that rehydration is not currently possible.

Files

dataset_tweet_index.csv.gz

Lightweight index of all 12,901,991 tweets in the corpus (254 MB, gzip-compressed).

tweet_id — Twitter status IDuser_id_anon — Anonymised user identifier (12-character SHA-256 hash)subset — Conference period: cop26 or cop27community — Community label from Infomap community detection (blank if user not in the network analysis)community_category — Community-level classification: disinfo, mixed, or non-disinfodisinfo_probability — Classifier-assigned probability that the tweet contains climate misinformation (0–1)is_disinfo — Binary label: 1 if disinfo_probability ≥ 0.5, else 0

Of the 12.9 million tweets, 8,614,156 are from COP26 and 4,287,835 from COP27. Community labels are available for the 674,485 tweets from the 19,003 users included in the network analysis.

dataset_network_tweets.csv

Full detail for the 674,485 tweets from users in the network analysis (88.5 MB).

tweet_id — Twitter status IDuser_id_anon — Anonymised user identifiersubset — Conference periodcreated_at — Tweet timestamp (original UTC format)lang — Tweet language codereply_count — Number of repliesretweet_count — Number of retweetsquote_count — Number of quote tweetslike_count — Number of likesretweeted_user_anon — Anonymised ID of retweeted user (blank if not a retweet)quoted_user_anon — Anonymised ID of quoted user (blank if not a quote tweet)replied_user_anon — Anonymised ID of replied-to user (blank if not a reply)community — Community labelcommunity_category — Community-level classificationdisinfo_probability — Classifier-assigned misinformation probabilityis_disinfo — Binary misinformation label

dataset_community_summary.csv

Aggregated statistics for each community (68 KB, 636 rows).

community — Community labelcommunity_category — Classification: disinfo, mixed, or non-disinfon_users — Number of unique usersn_tweets — Number of tweetsmean_disinfo_prob — Mean misinformation probabilitymedian_disinfo_prob — Median misinformation probabilitypct_disinfo — Proportion of tweets classified as misinformationmean_retweet_count — Mean retweet countmean_like_count — Mean like countmean_reply_count — Mean reply count

Method summary

Tweets were collected using the Twitter Academic Research API (v2). Misinformation classification uses a stacking ensemble (logistic regression, random forest, and support vector classifier with a gradient boosting meta-learner) trained on the CLIMATE-FEVER and GW Stance datasets. Community detection was performed using the Infomap algorithm on the retweet and quote-tweet interaction network. Communities were categorised as `disinfo`, `mixed`, or `non-disinfo` based on their aggregate misinformation probability distribution. Full methodological details are provided in the accompanying article.

Anonymisation

- All usernames are replaced with 12-character truncated SHA-256 hashes- Tweet text is not included in any file- Interaction targets (retweeted, quoted, replied-to users) are anonymised using the same hashing scheme- Tweet IDs are retained as these are public identifiers

Citation

If you use this dataset, please cite the accompanying article:

[citation to be added upon publication]