cy-330/classifai-screen-time

ClassifAI Screen-Time Screenshots: Study vs. Entertainment

Screenshot dataset from the UCSB ERSP 2026 project "Lightweight AI for Screen Time Classification" (Yan 2 group). The project builds ClassifAI, a privacy-preserving iPadOS app that classifies on-device screen content as study/work or entertainment using lightweight MobileNetV3 models.

• Code: oscarv06-cs/Yan-2
• Authors: Yue Cao, Rohan Iyer, Leeor Shemesh, Oscar Valeriano
• Technical mentor: Prof. Xifeng Yan

Configurations

Config	Split	Videos	Images	entertainment	study
training	train	222	35,637	17,461	18,176
benchmark	test	89	6,173	3,351	2,822

The benchmark is a held-out, source-disjoint set. None of its source videos appear in the training set (0 shared YouTube IDs), so it measures generalization to new content.

from datasets import load_dataset

train = load_dataset("cy-330/classifai-screen-time", "training", split="train")
bench = load_dataset("cy-330/classifai-screen-time", "benchmark", split="test")

print(train.features["label"].names)  # ['entertainment', 'study']

How the data was collected

1. We selected YouTube videos that show iPad or screen-recorded activity. Study videos include note-taking apps, coding, office tools, lectures and math problem-solving. Entertainment videos include mobile games, streaming, short-form social media, sports and vlogs. See screenshot_collection/video_url.csv (training) and benchmark/metadata/benchmark_seed_links_novel.csv (benchmark) in the code repo.
2. We extracted one frame every 10 seconds, capped at the first hour of each video (screenshot_collection/get_screenshot.py).
3. We removed near-duplicate frames within each video using dHash (screenshot_collection/dedup_frames_dhash.py).
4. Each frame inherits the label of its source video.

Fields

Field	Type	Description
image	Image (JPEG)	Video frame / screenshot
label	ClassLabel	0 = entertainment, 1 = study (same encoding as the repo CSVs)
video_id	string	YouTube video ID of the source video
video_name	string	Original frame-folder name (<title>_<video_id>)
video_title	string	Source video title
video_url	string	Source video URL
content_group	string	On-screen content group derived from the title, e.g. Note-taking apps or Mobile & casual games (rules in visualization/common.py::classify_content)
frame_index	int32	Frame index within the video
timestamp_s	int32	Frame timestamp in seconds. Null for 21,856 training frames (127 videos) whose filenames did not record it

Content coverage

Training (training):

label	content_group	videos	images
entertainment	Mobile & casual games	54	8,829
entertainment	Funny · reactions · challenges	14	2,673
entertainment	Vlogs · lifestyle · relaxation	9	1,732
entertainment	Streaming · movies · TV · anime	10	1,646
entertainment	Sports broadcasts	4	1,081
entertainment	Short-form & social media	11	912
entertainment	Esports	1	329
entertainment	Music & audio	4	176
entertainment	Shopping apps	2	69
entertainment	Other entertainment	1	14
study	Recorded lectures & courses	21	5,439
study	Coding & dev tools	17	3,837
study	Office & document tools	20	3,031
study	Note-taking apps	24	2,310
study	Math & STEM problem-solving	16	1,605
study	Study productivity & AI	3	578
study	Creative & design tools	3	539
study	Language learning	3	496
study	Reading & research	4	274
study	Other study	1	67

Benchmark (benchmark):

label	content_group	videos	images
entertainment	Short-form & social media	10	815
entertainment	Mobile & casual games	8	715
entertainment	Streaming · movies · TV · anime	10	677
entertainment	Esports	3	464
entertainment	Funny · reactions · challenges	3	347
entertainment	Vlogs · lifestyle · relaxation	3	177
entertainment	Sports broadcasts	1	81
entertainment	Music & audio	4	66
entertainment	Shopping apps	1	9
study	Recorded lectures & courses	9	616
study	Coding & dev tools	7	578
study	Math & STEM problem-solving	11	553
study	Office & document tools	4	298
study	Note-taking apps	4	189
study	Study productivity & AI	3	184
study	Creative & design tools	2	173
study	Reading & research	3	145
study	Language learning	3	86

Train / validation / test splits

Only one train split is published for the training data. The repo's screenshot_collection/split_train_test.py builds a roughly 80/10/10 train/val/test split at training time, after deduplication. Frames from the same video are highly correlated, so split by video_id if you need your own validation set.

Results (from the project)

The Model B work-vs-entertainment classifier (MobileNetV3-Small, 224×224 input) was evaluated on the benchmark config:

Model / setting	Accuracy	Macro-F1
causal_r2_interleave5 (soft-label model + causal temporal smoothing)	85.9%	0.857
rico1_soft_label	85.1%	0.849
rico, best grid setting (neighbor range 1, interleave block 20)	87.0%	0.868

Full per-video error analysis is in visualization/report_figures/REPORT.md in the code repo.

Limitations and biases

• Labels are assigned per video, not per frame. Frames such as intros, ads or home screens inside a video inherit the video's label, which adds label noise. The project's soft_label/ pipeline rescored ambiguous frames with a vision LLM to mitigate this.
• The dataset is based on YouTube recordings, so it over-represents polished tutorial-style content compared with real personal device use.
• Content is mostly English-language and skews toward iPad apps.

License and intended use

The images are frames from third-party YouTube videos. Copyright stays with the original creators, listed by video_url and video_title for attribution. This dataset is shared for non-commercial research and educational use only. If you are a rights holder and want content removed, please open a discussion on this repository.

Citation

@misc{cao2026classifai,
  title  = {Lightweight AI for Screen Time Classification},
  author = {Cao, Yue and Iyer, Rohan and Shemesh, Leeor and Valeriano, Oscar},
  year   = {2026},
  note   = {UCSB Early Research Scholars Program (ERSP), mentor: Xifeng Yan},
  howpublished = {\url{https://github.com/oscarv06-cs/Yan-2}}
}