The-data-company/TikTok-Longitudinal-Engagement

TikTok Longitudinal Engagement

The dataset contains repeated measurements of plays, likes, comments, shares, and saves for 99.5 million TikTok videos. These measurements support comparisons of engagement growth across videos of different ages and within the same videos across consecutive calendar periods.

Configuration	Rows	Unit
items	99,457,493	Video
item_stats	565,071,383	Recorded engagement observation
study_windows	10,356,618	Video and study window definition

The download totals 42.26 GiB. Each table has 256 Parquet files. Matching part numbers use the same video ID partition. A study_windows shard contains the videos in that partition with a qualifying window. The Hub exposes each full table under the train label.

Coverage and sampling

Recorded observations span March 7 to July 3, 2025 (UTC). Videos were observed through TikTok hashtag lists. A video is included if its first and last recorded observations are at least 24 hours apart. Each included video has its full available history, with cumulative counter values preserved as recorded.

Load a table

The three table names are Hugging Face dataset configurations. Streaming reads records as needed.

from datasets import load_dataset

items = load_dataset(
    "The-data-company/TikTok-Longitudinal-Engagement", "items", split="train", streaming=True
)
print(next(iter(items)))

Tables and joins

Join item_stats.item_id to items.id. item_stats.id identifies a single observation. Use the count and time range on items to filter histories before reading the statistics files.

Each study window has two endpoint UUIDs, start_stat_id and end_stat_id, which reference item_stats.id for the same video. The endpoint timestamps use an inclusive ±12-hour recorded-time matching tolerance. Ties favor the earlier recorded timestamp, then the lowest source UUID.

Definition	Rows	Endpoints
monthly_30d	9,530,800	Earliest qualifying recorded baseline and nearest observation to baseline plus 30 days
calendar_2025_04_01_05_01	149,493	Nearest observations to April 1 and May 1, 2025 at midnight UTC
calendar_2025_05_01_05_31	676,325	Nearest observations to May 1 and May 31, 2025 at midnight UTC

For a paired April–May comparison, select the 61,289 videos with both calendar windows. Each pair shares the same May 1 observation. Both video age and calendar conditions change between intervals.

Read one matching shard

Download part-000.parquet from both study_windows/ and item_stats/ and keep those directory names. The DuckDB query returns net play changes for ten windows in that partition.

SELECT w.item_id, w.definition, s.recorded_at AS start_at,
       e.recorded_at AS end_at, e.play_count - s.play_count AS net_plays
FROM read_parquet('study_windows/part-000.parquet') w
JOIN read_parquet('item_stats/part-000.parquet') s
  ON s.id = w.start_stat_id AND s.item_id = w.item_id
JOIN read_parquet('item_stats/part-000.parquet') e
  ON e.id = w.end_stat_id AND e.item_id = w.item_id
LIMIT 10;

manifest.json lists row counts, file sizes, and SHA-256 checksums.

Limitations

Inclusion depends on which videos appeared in the captured hashtag lists and when they were revisited. Comparisons describe the eligible videos in this collection. Observation intervals vary across videos, and cumulative counters can repeat or decrease.

create_time is publication time in Unix seconds. recorded_at is when the observation was written to the database, in UTC. The ±12-hour window tolerance applies to recorded times.

Column reference

items

Column	Arrow type	Nullable	Source or meaning
id	string	No	Source video identifier
metadata_created_at	timestamp[us, UTC]	No	Time the video metadata was first recorded
metadata_updated_at	timestamp[us, UTC]	No	Time the video metadata was last updated
author_tt_user_id	string	No	Source creator identifier
tt_video_id	string	No	Source video entity identifier
tt_poi_id	string	Yes	Source place identifier
tt_music_id	string	Yes	Source sound identifier
create_time	int64	No	TikTok publication time in Unix seconds
desc	string	No	Source description
duet_display	int16	No	Source duet display code
duet_enabled	bool	No	Source duet flag
duet_info_duet_from_id	string	Yes	Source duet parent video identifier
is_ad	bool	No	Source ad flag
item_comment_status	int16	No	Source comment status code
item_mute	bool	Yes	Source mute flag
item_control_can_repost	bool	No	Source repost permission flag
official_item	bool	No	Source official item flag
original_item	bool	No	Source original item flag
share_enabled	bool	No	Source sharing flag
text_language	string	Yes	Source language label
playlist_id	string	Yes	Source playlist identifier
stitch_display	int16	No	Source stitch display code
stitch_enabled	bool	No	Source stitch flag
diversification_id	int16	Yes	Source diversification code
observation_count	int64	No	Number of released statistics rows for the video
distinct_recorded_times	int64	No	Number of distinct statistics timestamps
first_recorded_at	timestamp[us, UTC]	No	Earliest released statistics timestamp
last_recorded_at	timestamp[us, UTC]	No	Latest released statistics timestamp

item_stats

Column	Arrow type	Nullable	Source or meaning
id	string	No	Original statistics observation UUID
recorded_at	timestamp[us, UTC]	No	Time the observation was written to the database
item_id	string	No	Video ID, references items.id
collect_count	int64	No	Saves or collects counter
comment_count	int64	No	Comments counter
digg_count	int64	No	Likes counter
play_count	int64	No	Plays counter
share_count	int64	No	Shares counter

study_windows

Column	Arrow type	Nullable	Meaning
item_id	string	No	References items.id
definition	string	No	Window rule identifier
start_stat_id	string	No	Starting observation UUID, references item_stats.id
end_stat_id	string	No	Ending observation UUID, references item_stats.id

Citation

@misc{holt2026tiktokengagement,
  author       = {Holt, Andrew},
  title        = {{TikTok Longitudinal Engagement}},
  year         = {2026},
  howpublished = {The Data Company},
  url          = {https://huggingface.co/datasets/The-data-company/TikTok-Longitudinal-Engagement}
}