Skip to main content

WatchLens: A Configurable Platform for Online Video Recommendation Experiments

Deogyong Kim, Dongha Lee*

Department of Artificial Intelligence, Yonsei University

*Corresponding author

PaperarXivCodeDocs
Pluggable Recommenders

Pluggable Recommenders

Add a recommendation policy as a single Python class, or register an externally hosted HTTP service through the admin API. Five built-ins ship (random, popularity, recency, similarity, RecBole), all behind one dispatcher and one capability contract.

Standardized Event Schema

Standardized Event Schema

33 events across 6 categories (session, navigation, playback, impressions, interactions, browser state) emit automatically from every UI track. Recommendation policy and feed position attach to every row, so cross-treatment comparison is the default.

Configurable UI

Configurable UI

Three authoring tracks share the same event contract. The bundled YouTube and TikTok presets, an in-browser TSX code editor compiled with sucrase, and a visual block-tree editor with 19 composable blocks. Eject from visual to code when you outgrow the library.

BibTeX

@inproceedings{kim2026watchlens,
  title     = {WatchLens: A Configurable Platform for Online Video Recommendation Experiments},
  author    = {Kim, Deogyong and Lee, Dongha},
  booktitle = {Proceedings of the 20th ACM Conference on Recommender Systems},
  series    = {RecSys '26},
  pages     = {927--932},
  year      = {2026},
  publisher = {Association for Computing Machinery},
  doi       = {10.1145/3773078.3831858}
}