KuaiRand-Star

Star-schema Parquet for the official KuaiRand releases. pure, 1k, and 27k are separate subsets. Each subset has its own events, users, items, item_statistics, and item_tags tables. The tables are not joined. 27k is the full release. 1k samples 1,000 users from 27k and drops videos those users did not meet, so its side tables are a different release file. pure keeps the logs of the 7,583 candidate-pool videos. video_captions and video_categories are two shared configs from the supplementary Zenodo record. They are not split by subset. Field meanings below follow the descriptions on that homepage.

This is not the official repository. A config name is {subset}-{table}. The default config is 1k-events.

Rows

subset	events	users	items	item_statistics	item_tags
pure	2,622,668	27,285	7,583	7,583	9,310
1k	11,756,073	1,000	4,371,868	4,371,868	5,207,822
27k	323,464,444	27,285	32,038,725	32,038,725	38,902,423

The events count is the official standard-policy interactions plus the random-policy interactions: pure 1,436,609 + 1,186,059, 1k 11,713,045 + 43,028, and 27k 322,278,385 + 1,186,059. pure and 27k each have 27,285 users, and their user_features CSVs are byte-identical. 1k has 1,000 users.

The official README counts 4,369,953 videos in the 1k standard-policy logs and 7,388 in the random log. 1k items has 4,371,868 rows, the number of distinct video_id values across both logs. 27k items has the official 32,038,725 videos. pure items has 7,583 rows, the candidate-pool size in the official random-policy row.

events.event_at on pure runs from 2022-04-09 03:46:03.278 through 2022-05-08 15:52:10.564. On 1k it runs from 2022-04-07 14:00:03.236 through 2022-05-08 15:52:09.022. On 27k it runs from 2022-04-07 13:59:26.592 through 2022-05-08 15:52:10.589.

Columns

Every original CSV field is kept as VARCHAR. Empty source fields stay null. They are not filled with zero.

On events, item_id is video_id as BIGINT and event_at is epoch_ms(time_ms). user_id stays the original decimal string because that is already the CSV column name. date, hourmin, and is_rand stay as released. Standard logs have is_rand 0 and the random log has is_rand 1. Rows that share user_id and event_at are kept.

On items and item_statistics, item_id is video_id as BIGINT. author_id stays author_id. items.tag stays the original comma-separated string. item_tags is that string split on comma, trimmed, with empty tokens omitted. position is the 1-based index in the original string. item_tags.source_row is the parent item row, so it is not unique. users, items, and item_statistics have no clock.

source_row on events, users, items, and item_statistics is the 1-based row number after concatenating the official files and before sorting the Parquet file. Log order is standard before random, then the 4_08_to_4_21 span before 4_22_to_5_08, then part number. Statistic parts concatenate in part order. The written events file is sorted by user_id, event_at, source_row.

dataset_source is kuairand and synthetic is false on every row. subset is pure, 1k, or 27k on the subset tables. The two shared configs have source_row, dataset_source, and synthetic, and they have no subset column. item_id on those configs is final_video_id as BIGINT.

Field meanings

The homepage describes three log files. log_standard_4_08_to_4_21 is standard recommendation for 2022-04-08 through 2022-04-21. log_standard_4_22_to_5_08 is standard recommendation in the later span. log_random_4_22_to_5_08 is the random intervention. The homepage says each recommended video had an equal chance of being replaced by a video drawn from the candidate pool, and that about 0.37% of interactions are those replacements. is_rand is 1 on a replaced row. tab is the scenario and takes an integer from 0 through 14, one value for each of the 15 policies.

field	meaning on events
user_id	User id. The same id space as users.
video_id	Video id. The same id space as items.
date	Interaction date as YYYYMMDD, for example 20220421.
hourmin	Interaction time in the homepage's HHSS form, for example 400.
time_ms	Interaction timestamp in milliseconds. event_at is epoch_ms of this value.
is_click	Binary feedback. On the two-column UI it is a click. On the single-column UI it is a valid play: 1 when play_time_ms >= duration_ms if duration_ms <= 7000, and when play_time_ms > 7000 if duration_ms > 7000.
is_like	The user hit like.
is_follow	The user followed the author.
is_comment	The user wrote a comment.
is_forward	The user forwarded the video.
is_hate	The user marked the video as hated.
long_view	1 when play_time_ms >= duration_ms if duration_ms <= 18000, and when play_time_ms >= 18000 if duration_ms > 18000.
play_time_ms	View time in milliseconds.
duration_ms	Video duration in milliseconds.
profile_stay_time	Time spent on the author profile.
comment_stay_time	Time spent in the comments.
is_profile_enter	The user entered the author profile.
is_rand	1 when this impression is a random exposure.
tab	Scenario id, 0 through 14.

users has 30 features plus user_id. user_active_degree is one of high_active, full_active, middle_active, or UNKNOWN. is_lowactive_period, is_live_streamer, and is_video_author are binary. follow_user_num, fans_user_num, and friend_user_num are counts, and the matching *_range columns are the released range labels. register_days is days since registration. Its range is one of 15-30, 31-60, 61-90, 91-180, 181-365, 366-730, or 730+. onehot_feat0 through onehot_feat17 are encrypted features. Each value is the position of the 1 in a one-hot vector. The homepage gives these ranges: feat0 is 0 or 1; feat1 is 0 through 6; feat2 is 0 through 49; feat3 is 0 through 1470; feat4 is 0 through 14; feat5 is 0 through 33; feat6 is 0, 1, or 2; feat7 is 0 through 117; feat8 is 0 through 453; feat9 is 0 through 6; feat10 and feat11 are 0 through 4; feat12 through feat17 are 0 or 1.

field	meaning on items
author_id	Author id. The homepage gives the range 0 through 8,839,734.
video_type	NORMAL or AD.
upload_dt	Upload date, for example 2020-07-08.
upload_type	How the video was uploaded, for example ShortImport.
visible_status	Visibility on the app in the released snapshot.
video_duration	Duration in milliseconds. The released text can look like 17200.0.
server_width, server_height	Frame size on the server.
music_id, music_type	Background music id and type.
tag	Comma-separated category labels, for example 12,65. item_tags is this list unnested.

item_statistics is one row per video. The homepage says each value is the average of that statistic over each day and each scenario in one month. counts is the number of observations in that average. A video can contribute more than one observation on the same day when it is counted on more than one scenario. The table has no date, so it is not a changelog and it has no valid_from.

On that table, show_cnt is shows and show_user_num is users who received the video. play_cnt, play_user_num, and play_duration are plays, playing users, and total play time in milliseconds. play_progress is the average of play_duration / video_duration. A complete play finishes the video (play_duration >= video_duration). A valid play is play_duration >= video_duration when the video is at most 7 seconds, and play_duration > 7 seconds otherwise. A long-time play uses 18 seconds in that same pattern. A short-time play is play_duration < min(3 seconds, video_duration). Each of those four has a *_cnt and a *_user_num. The remaining pairs are likes, double-click likes (click_like_cnt, double_click_cnt), canceled likes, comments, direct comments, reply comments, deleted comments, comment likes, follows, canceled follows, successful shares, all share attempts (share_all_cnt), shares outside Kuaishou (outsite_share_all_cnt), downloads, reports, reduce-similar, collects, and canceled collects. comment_stay_duration is time in the comments. follow_cnt counts new follows attributed to the video.

Video captions and categories

These two configs are the supplementary files at https://zenodo.org/records/18159199. Each file is one table. The rows are not copied into pure, 1k, or 27k, and they are not joined to items. final_video_id stays the original text. The homepage says it is the re-indexed KuaiRand video id. source_row is the 1-based CSV row before sorting by final_video_id, source_row.

video_captions has 32,038,725 rows and 32,038,725 distinct final_video_id values. The intersection with items.video_id is 7,583 of the 7,583 pure video_id values, 4,371,868 of the 4,371,868 1k video_id values, and 32,038,725 of the 32,038,725 27k video_id values. That is the same id set as 27k items.video_id. Every video_id in pure and 1k is in this file.

field	meaning on video_captions
final_video_id	Re-indexed video id. The homepage types it as int64.
caption	Caption text, including hashtags and mentions.
show_cover_text	Text on the video cover. It may be empty.
duration	Duration in milliseconds. The released text keeps the fractional form, for example 114500.0.

video_categories has 32,038,725 rows and 32,038,725 distinct final_video_id values. The intersection with items.video_id is 7,583 of the 7,583 pure video_id values, 4,371,868 of the 4,371,868 1k video_id values, and 32,038,725 of the 32,038,725 27k video_id values. That is the same id set as 27k items.video_id. Every video_id in pure and 1k is in this file.

video_categories is a four-level category prediction. Each level is already a column: an id, a name, and a probability. The levels are not unnested. A missing level uses id -124 and name UNKNOWN in the homepage description. The id and probability columns stay the original CSV text, including a trailing .0 when the file has one.

field	meaning on video_categories
first_level_category_id, first_level_category_name, first_level_category_prob	Top category id, name, and predicted probability.
second_level_category_id, second_level_category_name, second_level_category_prob	Second-level category.
third_level_category_id, third_level_category_name, third_level_category_prob	Third-level category, when the file has one.
fourth_level_category_id, fourth_level_category_name, fourth_level_category_prob	Fourth-level category. The homepage example of a missing level is id -124.0 and name UNKNOWN.

Source files

• pure events: log_standard_4_08_to_4_21_pure.csv (1,141,112), log_standard_4_22_to_5_08_pure.csv (295,497), log_random_4_22_to_5_08_pure.csv (1,186,059)
• pure users: user_features_pure.csv (27,285)
• pure items: video_features_basic_pure.csv (7,583)
• pure item_statistics: video_features_statistic_pure.csv (7,583)
• 1k events: log_standard_4_08_to_4_21_1k.csv (5,055,984), log_standard_4_22_to_5_08_1k.csv (6,657,061), log_random_4_22_to_5_08_1k.csv (43,028)
• 1k users: user_features_1k.csv (1,000)
• 1k items: video_features_basic_1k.csv (4,371,868)
• 1k item_statistics: video_features_statistic_1k.csv (4,371,868)
• 27k events: log_standard_4_08_to_4_21_27k_part1.csv (68,148,288), log_standard_4_08_to_4_21_27k_part2.csv (68,148,288), log_standard_4_22_to_5_08_27k_part1.csv (93,062,020), log_standard_4_22_to_5_08_27k_part2.csv (92,919,789), log_random_4_22_to_5_08_27k.csv (1,186,059)
• 27k users: user_features_27k.csv (27,285)
• 27k items: video_features_basic_27k.csv (32,038,725)
• 27k item_statistics: video_features_statistic_27k_part1.csv (10,679,575), video_features_statistic_27k_part2.csv (10,679,575), video_features_statistic_27k_part3.csv (10,679,575)
• video_captions: kuairand_video_captions.csv (32,038,725)
• video_categories: kuairand_video_categories.csv (32,038,725)

Original sources

• Homepage: https://kuairand.com/
• GitHub: https://github.com/chongminggao/KuaiRand
• Zenodo: https://zenodo.org/records/10439422
• Video captions and categories: https://zenodo.org/records/18159199
• Paper: https://doi.org/10.1145/3511808.3557624

License

CC BY-SA 4.0, the license of the official KuaiRand release.

Citation

@inproceedings{gao2022kuairand,
  author = {Gao, Chongming and Li, Shijun and Zhang, Yuan and Chen, Jiawei and Li, Biao and Lei, Wenqiang and Jiang, Peng and He, Xiangnan},
  title = {KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos},
  booktitle = {Proceedings of the 31st ACM International Conference on Information and Knowledge Management},
  series = {CIKM '22},
  year = {2022},
  pages = {3953--3957},
  doi = {10.1145/3511808.3557624}
}