KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos

KuaiRand is an unbiased sequential recommendation dataset collected from the recommendation logs of the video-sharing mobile app, Kuaishou (快手). It is the first recommendation dataset with millions of intervened interactions of randomly exposed items inserted in the standard recommendation feeds!

Overview:

The following figure gives an example of the dataset. It illustrates a user interaction sequence along with the user's rich feedback signals.

These feedback signals are collected from the two main user interfaces (UI) in Kuaishou APP shown as follows:

In the random exposure stage, each recommended video in the dataset has an equal probability being replaced by a random video sampled from an item pools. About $0.37%$ Interactions are replaced in the final results.

Advantages:

Compared with other datasets with random exposure, KuaiRand has the following advantages:

• ✅ It is the first sequential recommendation dataset with millions of intervened interactions of randomly exposed items inserted in the standard recommendation feeds.
• ✅ It has the most comprehensive side information including explicit user IDs, interaction timestamps, and rich features for users and items.
• ✅ It has 15 policies with each catered for a special recommendation scenario in Kuaishou App.
• ✅ We introduce 12 feedback signals (e.g., click, like, and view time) for each interaction to describe the user's comprehensive feedback.
• ✅ Each user has thousands of historical interactions on average.
• ✅ It has three versions to support various research directions in recommendation.

---

Three Versions and Suggestions:

We release three versions of KuaiRand for different usage:

• KuaiRand-27K (23GB logs +23GB features): the complete KuaiRand dataset that has over 27K users and 32 millions videos.
• KuaiRand-1K (829MB logs + 3.5GB features): randomly sample 1,000 users from KuaiRand-27K, then remove all irrelevant videos. There are 4 millions videos rest.
• KuaiRand-Pure (184MB logs + 10MB features): only keeps the logs for the 7582 videos in the candidate pool.

The user_id and video_id are re-indexed. A visualization of their ID spaces is shown as follows.

The basic statistics of the three versions are summarized as follows:

Dataset	Collection Policy	#Users	#Items	#Interactions	#User features	#Item features	Feedback	Timestamp
KuaiRand-27K	15 policies	27,285	32,038,725	322,278,385	30	62	12 signals	✔️
	Random policy	27,285	7,583	1,186,059	30	62	12 signals	✔️
KuaiRand-1K	15 policies	1,000	4,369,953	11,713,045	30	62	12 signals	✔️
	Random policy	1,000	7,388	43,028	30	62	12 signals	✔️
KuaiRand-Pure	15 policies	27,285	7,551	1,436,609	30	62	12 signals	✔️
	Random policy	27,285	7,583	1,186,059	30	62	12 signals	✔️

Which version should I use?

• Reasons to use KuaiRand-27K or KuaiRand-1K:

  • Your research needs rigorous sequential logs, such as off-policy evaluation (OPE), Reinforcement learning (RL), or long sequential recommendation.

• Reasons to use KuaiRand-Pure:

  • The sequential information is not necessary in your research OR If you are OK with the incomplete sequential logs. For example, if you are studying debiasing in collaborative filtering models or multi-task modeling in recommendation.
  • If your model can only run with small-size data.

---

Download the data

We provides several options to download the datasets:

Option 1. Download via the "wget" command.

KuaiRand-27K (compressed size: 9.21GB; uncompressed size: 46GB)

wget https://chongming.myds.me:61364/data/KuaiRand-27K.tar.gz --no-check-certificate
tar -zxvf KuaiRand-27K.tar.gz

KuaiRand-1K (compressed size: 1GB; uncompressed size: 4.3GB)

wget https://chongming.myds.me:61364/data/KuaiRand-1K.tar.gz --no-check-certificate
tar -zxvf KuaiRand-1K.tar.gz

KuaiRand-Pure (compressed size: 45.2MB; uncompressed size: 194MB)

wget https://chongming.myds.me:61364/data/KuaiRand-Pure.tar.gz --no-check-certificate
tar -zxvf KuaiRand-Pure.tar.gz

Option 2. Download manually through the following links:

• Optional link 1: Google Drive

• Optional link 2: USTC Drive (中科大)

The script loaddata.py provides a simple way to load the data via Pandas in Python.

---

Data Descriptions

The file structure of the three datasets are listed as follows:

KuaiRand-27K (46GB)

KuaiRand-27K
├── data (46GB)
│   ├── log_random_4_22_to_5_08_27k.csv (83MB)
│   ├── log_standard_4_08_to_4_21_27k_part1.csv (4.8GB)
│   ├── log_standard_4_08_to_4_21_27k_part2.csv (4.8GB)
│   ├── log_standard_4_22_to_5_08_27k_part1.csv (6.6GB)
│   ├── log_standard_4_22_to_5_08_27k_part2.csv (6.6GB)
│   ├── user_features_27k.csv (3.4MB)
│   ├── video_features_basic_27k.csv (2.6GB)
│   ├── video_features_statistic_27k_part1.csv (6.7GB)
│   ├── video_features_statistic_27k_part2.csv (6.7GB)
│   └── video_features_statistic_27k_part3.csv (6.7GB)
└── load_data_27k.py

KuaiRand-1K (4.3GB)

KuaiRand-1K
├── data (4.3GB)
│   ├── log_random_4_22_to_5_08_1k.csv (2.9MB)
│   ├── log_standard_4_08_to_4_21_1k.csv (368MB)
│   ├── log_standard_4_22_to_5_08_1k.csv (481MB)
│   ├── user_features_1k.csv (132KB)
│   ├── video_features_basic_1k.csv (368MB)
│   └── video_features_statistic_1k.csv (3.1GB)
└── load_data_1k.py

KuaiRand-Pure (194MB)

KuaiRand-Pure
├── data (194MB)
│   ├── log_random_4_22_to_5_08_pure.csv (83MB)
│   ├── log_standard_4_08_to_4_21_pure.csv (80MB)
│   ├── log_standard_4_22_to_5_08_pure.csv (21MB)
│   ├── user_features_pure.csv (3.4MB)
│   ├── video_features_basic_pure.csv (612KB)
│   └── video_features_statistic_pure.csv (6.3MB)
└── load_data_pure.py

---

1️⃣ Descriptions of the fields in log_xxx.csv

There are three log files:

• log_random_4_22_to_5_08.csv contains all interactions resulted from random intervention.
• log_standard_4_22_to_5_08.csv contains all interactions of standard recommendation.
• log_standard_4_08_to_4_21.csv contains all interactions of standard recommendation for the same users in the previous two weeks (2022.04.08 ~ 2022.04.21).

Field Name:	Description	Type	Example
user_id	The ID of the video.	int64	17387
video_id	The ID of the video.	int64	1123453
date	The date of this interaction	int64	20220421
hourmin	The time of this interaction (format: HHSS).	int64	400
time_ms	The timestamp of this interaction in millisecond.	int64	1650485801301
is_click	A binary feedback signal. In the two-column UI, it indicates a click; In the single-column UI, it means valid_play: which equals 1 when: play_duration >= video_duration if video_duration <= 7s, or play_duration > 7 if video_duration > 7s.	int64	1
is_like	A binary feedback signal indicating the user hit the like button.	int64	0
is_follow	A binary feedback signal indicating the user hit the follow the author button.	int64	0
is_comment	A binary feedback signal indicating the user writed comment in the comments section of this video	int64	0
is_forward	A binary feedback signal indicating the user forwarded this video.	int64	0
is_hate	A binary feedback signal indicating the user hated this video.	int64	0
long_view	A binary feedback signal. It equals 1 when: play_duration >= video_duration if video_duration <= 18s, or play_duration >=18 if video_duration > 18s.	int64	1
play_time_ms	The user's view time in millisecond.	int64	151024
duration_ms	The video's duration time in millisecond.	int64	104400
profile_stay_time	The time that the user stayed in this author's profile.	int64	0
comment_stay_time	The time that the user stayed in the comments section of this video	int64	0
is_profile_enter	A binary feedback signal indicating the user enters the author profile	int64	0
is_rand	A binary signal indicating if this video is generated by the random intervention (i.e., a random exposed video).	int64	0
tab	indicating the scenario of this interaction, e.g., in the recommendation page or the main page of the App. In range of [0,14].	int64	1

---

2️⃣ Descriptions of the fields in user_features.csv

Field Name:	Description	Type	Example
user_id	The ID of the user.	int64	25621
user_active_degree	In the set of {'high_active', 'full_active', 'middle_active', 'UNKNOWN'}.	str	"full_active"
is_lowactive_period	Is this user in its low active period	int64	0
is_live_streamer	Is this user a live streamer？	int64	0
is_video_author	Has this user uploaded any video？	int64	1
follow_user_num	The number of users that this user follows.	int64	5
follow_user_num_range	The range of the number of users that this user follows. In the set of {'0', '(0,10]', '(10,50]', '(100,150]', '(150,250]', '(250,500]', '(50,100]', '500+'}	str	"(0,10]"
fans_user_num	The number of the fans of this user.	int64	312
fans_user_num_range	The range of the number of fans of this user. In the set of {'0', '[1,10)', '[10,100)', '[100,1k)', '[1k,5k)', '[5k,1w)', '[1w,10w)'}	str	"[100,1k)"
friend_user_num	The number of friends that this user has.	int64	0
friend_user_num_range	The range of the number of friends that this user has. In the set of {'0', '[1,5)', '[5,30)', '[30,60)', '[60,120)', '[120,250)', '250+'}	str	"0"
register_days	The days since this user has registered.	int64	3624
register_days_range	The range of the registered days. In the set of {'15-30', '31-60', '61-90', '91-180', '181-365', '366-730', '730+'}.	str	"730+"
onehot_feat0	An encrypted feature of the user. Each value indicate the position of "1" in the one-hot vector. Range: {0,1}	int64	1
onehot_feat1	An encrypted feature. Range: {0, 1, ..., 6}	int64	2
onehot_feat2	An encrypted feature. Range: {0, 1, ..., 49}	int64	2
onehot_feat3	An encrypted feature. Range: {0, 1, ..., 1470}	int64	1153
onehot_feat4	An encrypted feature. Range: {0, 1, ..., 14}	int64	4
onehot_feat5	An encrypted feature. Range: {0, 1, ..., 33}	int64	0
onehot_feat6	An encrypted feature. Range: {0, 1, 2}	int64	0
onehot_feat7	An encrypted feature. Range: {0, 1, ..., 117}	int64	31
onehot_feat8	An encrypted feature. Range: {0, 1, ..., 453}	int64	354
onehot_feat9	An encrypted feature. Range: {0, 1, ..., 6}	int64	3
onehot_feat10	An encrypted feature. Range: {0, 1, ..., 4}	int64	3
onehot_feat11	An encrypted feature. Range: {0, 1, ..., 4}	int64	2
onehot_feat12	An encrypted feature. Range: {0, 1}	int64	1
onehot_feat13	An encrypted feature. Range: {0, 1}	int64	0
onehot_feat14	An encrypted feature. Range: {0, 1}	int64	0
onehot_feat15	An encrypted feature. Range: {0, 1}	int64	0
onehot_feat16	An encrypted feature. Range: {0, 1}	int64	0
onehot_feat17	An encrypted feature. Range: {0, 1}	int64	0

---

3️⃣ Descriptions of the fields in video_features_basic.csv

Field Name:	Description	Type	Example
video_id	The ID of the video.	int64	3784
author_id	The ID of the author of this video. In the range of [0, 8839734]	int64	441
video_type	Type of this video (NORMAL or AD).	str	"NORMAL"
upload_dt	Upload date of this video.	str	"2020-07-08"
upload_type	The upload type of this video.	str	"ShortImport"
visible_status	The visible state of this video on the APP now.	int	1
video_duration	The time duration of this duration (in millisecond).	Int64	17200.0
server_width	The width of this video on the server.	int64	720
server_height	The height of this video on the server.	int64	1280
music_id	Background music ID of this video.	int64	989206467
music_type	Background music type of this video.	int64	4
tag	A list of key categories (labels) of this video.	str	"12,65"

---

4️⃣ Descriptions of the fields in video_features_statistic.csv

‼️ Different from the basic features, the statistical features are the average statistics of the video each day over the one month. For example, in the following table, the video 9288071 have 66 counts over this one month (a video can have multiple counts each day on different scenarios, e.g., on April 8, show_cnt=80 on the main page and show_cnt=65 on the recommendation page of the App)

Field Name:	Description	Type	Example
video_id	The ID of the video.	int64	9288071
counts	The number of statistics.	int64	66
show_cnt	The number of shows of this video (averaged on each day and each scenario over the one month. This applies to all the following fields)	float64	75.212
show_user_num	The number of users who received the recommendation of this video.	float64	66.985
play_cnt	The number of plays.	float64	9.409
play_user_num	The number of users who plays this video.	float64	8.121
play_duration	The total time duration of playing this video (in millisecond).	float64	93700.333
complete_play_cnt	The number of complete plays. complete play: finishing playing the whole video, i.e., #(play_duration >= video_duration).	float64	0.182
complete_play_user_num	The number of users who perform the complete play.	float64	0.182
valid_play_cnt	valid play: play_duration >= video_duration if video_duration <= 7s, or play_duration > 7 if video_duration > 7s.	float64	3.545
valid_play_user_num	The number of users who perform the complete play.	float64	3.136
long_time_play_cnt	long time play: play_duration >= video_duration if video_duration <= 18s, or play_duration >=18 if video_duration > 18s.	float64	1.909
long_time_play_user_num	The number of users who perform the long time play.	float64	1.848
short_time_play_cnt	short time play: play_duration < min(3s, video_duration).	float64	5.015
short_time_play_user_num	The number of users who perform the short time play.	float64	4.545
play_progress	The average video playing ratio (=play_duration/video_duration)	float64	0.016
comment_stay_duration	Total time of staying in the comments section	float64	2302.712
like_cnt	Total likes	float64	0.303
like_user_num	The number of users who hit the "like" button.	float64	0.303
click_like_cnt	The number of the "like" resulted from double click	float64	0.030
double_click_cnt	The number of users who double click the video.	float64	0.273
cancel_like_cnt	The number of likes that are cancelled by users.	float64	0.485
cancel_like_user_num	The number of users who cancel their like.	float64	0.485
comment_cnt	The number of comments within this day.	float64	0.015
comment_user_num	The number of users who comment this video.	float64	0.015
direct_comment_cnt	The number of direct comments (depth=1).	float64	0.015
reply_comment_cnt	The number of reply comments (depth>1).	float64	0.000
delete_comment_cnt	The number of deleted comments.	float64	0.015
delete_comment_user_num	The number of users who delete their comments.	float64	0.015
comment_like_cnt	The number of comment likes.	float64	0.000
comment_like_user_num	The number of users who like the comments.	float64	0.000
follow_cnt	The number of increased follows from this video.	float64	0.000
follow_user_num	The number of users who follow the author of this video due to this video.	float64	0.000
cancel_follow_cnt	The number of decreased follows from this video.	float64	0.000
cancel_follow_user_num	The number of users who cancel their following of the author of this video due to this video.	float64	0.000
share_cnt	The times of successfully sharing this video.	float64	0.000
share_user_num	The number of users who succeed to share this video.	float64	0.000
download_cnt	The times of downloading this video.	float64	0.030
download_user_num	The number of users who download this video.	float64	0.030
report_cnt	The times of reporting this video.	float64	0.000
report_user_num	The number of users who report this video.	float64	0.000
reduce_similar_cnt	The times of reducing similar content of this video.	float64	0.015
reduce_similar_user_num	The number of users who choose to reduce similar content of this video.	float64	0.015
collect_cnt	The times of adding this video to favorite videos.	float64	0.061
collect_user_num	The number of users who add this video to their favorite videos.	float64	0.061
cancel_collect_cnt	The times of removing this video from favorite videos.	float64	0.091
cancel_collect_user_num	The number of users who remove this video from their favorite videos	float64	0.091
direct_comment_user_num	The number of users who write comments directly under this video (level=1).	float64	0.015
reply_comment_user_num	The number of users who reply the existing comments (level>1).	float64	0.000
share_all_cnt	The times of sharing this video (no need to be successful).	float64	0.015
share_all_user_num	The number of users who share this video (no need to be successful).	float64	0.015
outsite_share_all_cnt	The times of sharing this video outside Kuaishou App.	float64	0.000