Open dataset · Kuaishou

Live streaming recommendation, in motion

KuaiLive-M3

A Multi-Modal, Multi-Domain, and Multi-Feedback Dataset for Live Streaming Recommendation

KuaiLive-M3 connects temporally evolving live content, short-video behaviors, and questionnaire-based feedback in one large-scale, real-world benchmark.

Explore on Hugging Face Browse field descriptions Benchmark code

Dataset profilePUBLIC RELEASE

KuaiLiveM3Recommendation research dataset

Multi-domain2connected ecosystems

Multi-feedback25Kquestionnaire responses

21,938 users146M interactions18 data collections

Why KuaiLive-M3

Three dimensions.
One living ecosystem.

Designed to close three long-standing gaps in public live-streaming datasets: evolving multimodal content, connected cross-domain behavior, and explicit user feedback.

Multi-modal

Content that evolves with the room

Timestamped 128-dimensional segment embeddings capture how live content changes over time, complemented by room-level live representations and video-level short-video embeddings.

Multi-domain

One audience, two connected domains

Shared, anonymized users and creators connect detailed live-stream and short-video histories, enabling realistic cross-domain transfer and recommendation studies.

Multi-feedback

Behavior meets explicit judgment

Rich implicit actions—watching, liking, commenting, gifting, following, sharing—are paired with real questionnaire-based feedback from live-stream viewers.

At a glance

Scale for real-world questions

Fine-grained records across both domains, anchored by a shared user population and temporally aligned live content.

0121,938Users

0235MLive interactions

03111MShort-video interactions

0488MLive segment embeddings

056.56MLive rooms

066.74MShort videos

0725KQuestionnaire responses

0818Data tables & collections

Rounded counts are provided for readability. Refer to the released files for exact row counts.

Dataset structure

Organized around two domains and one shared identity space

IDs are anonymized and remapped to consecutive integers. Live-stream files preserve room lifecycles and millisecond-level behaviors; short-video files include both aggregate and event-level playback records.

Segment timing. Each segment embedding is paired with its end timestamp, allowing alignment with watch sessions and other timestamped events.

View all 18 tables and fields →

·author_profile.csvcreator profiles

·user_id_set.csvanonymized user universe

▸live/live-streaming domain

·live_interaction.csvwatch sessions & actions

·live_show.parquetfeed impressions

·live_comment.csvtimestamped comments

·live_like.csv · live_share.csvfine-grained events

·live_questionnaire.csvexplicit feedback

·live_room_meta.parquetroom lifecycle & metadata

·live_emb_64.parquetroom-level embeddings

▸live_emb_128_ts/18 segment-embedding shards

▸short_video/short-video domain

·photo_interaction.csvdaily behavior aggregates

·photo_play.parquetevent-level playback

·photo_meta.parquet · photo_tag.csvmetadata & taxonomy

·photo_emb_128.parquetvideo-level embeddings

Official benchmark

Three tasks for the three dimensions of M3

The benchmark repository provides task-specific preprocessing, standardized evaluation protocols, representative baselines, and reproducible experiment runners.

View benchmark on GitHub ↗

01Multi-domain

Cross-domain recommendation

Transfer preferences from short-video viewing to live-stream author recommendation over a shared user population.

Protocol: Target-domain chronological 80/10/10 split; source domain for training
Metrics: Recall and NDCG @ 10, 20, 40
Models: BPR, SASRec, LightGCN + 12 CDR baselines

CMFCoNetEMCDRDisenCDRMGCCDR+7

02Multi-modal

Live-stream highlight prediction

Rank engaging segments using timestamped content embeddings and causally aligned retention and engagement signals.

Protocol: Room-level chronological 70/10/20 split; online and offline settings
Metrics: mAP, F1@50%, Kendall's tau, Spearman's rho
Inputs: Embedding, causal statistics, and embedding + statistics

MLPGRUCausal TransformerHierarchical Transformer

03Multi-feedback

Questionnaire-based recommendation

Complement click histories with satisfied and dissatisfied questionnaire signals while preventing answer leakage.

Protocol: Author-level 5-core with chronological leave-one-out evaluation
Metrics: MRR, HR, and NDCG over 1 positive + 99 negatives
Models: Click-only, multi-behavior, and satisfaction-aware methods

SASRecFeedRecDFNDMTSAQRec+8

Open access

Start exploring KuaiLive-M3

Dataset files are hosted on Hugging Face. Field-level documentation is available on this site.

Open dataset ↗Benchmark code ↗