desBugger/youtubeAlignment

YouTube Alignment: Recommendation Title Embeddings

Dataset accompanying the paper "Evaluating feedback mechanisms for aligning YouTube recommendations with user interests" (Cho, Hale, Zhao, Shadbolt & Przybylski; University of Oxford).

Dataset Summary

This dataset contains YouTube homepage recommendation titles and their semantic embeddings, collected from 12 simulated adolescent puzzle-gamer accounts across four experimental personas. It supports a controlled output-based audit of how YouTube's explicit feedback controls ('like' and 'not-interested') shift recommendation content relative to implicit engagement alone.

Experimental Design

We used a 2x2 between-subjects design crossing two explicit feedback conditions: (i) liking watched puzzle-gaming videos and (ii) clicking 'not-interested' on non-puzzle-gaming homepage recommendations. This yielded four personas: Watch (control), Like, Not-Interested, and Combined.

Three independent accounts were created per persona (12 accounts total). Accounts were modelled on a 13-year-old UK adolescent puzzle-gamer archetype (DoB: 12 March 2012; gender: Rather not say). All accounts were trained in synchrony over 7 days on a randomised playlist of 70 puzzle-gaming videos (10 videos/day). Homepage recommendations were scraped hourly during a 5-day maintenance phase (days 3-7). The 2-day warm-up period was excluded from analysis. Device and software were fixed across all accounts (MacBook Pro 13", macOS Ventura 13.0.1, Chrome 134.0.6998.89; single UK geolocation) to minimise confounds.

Data Collection

Recommendation titles were collected using an automated scraper adapted from TheirTube (Kihara 2020). Standard video recommendations were retained; Shorts and advertisements were excluded. Titles were cleaned by removing emojis, punctuation, and excess whitespace, and only English-language titles were kept.

Total recommendations collected: Watch (control) 1,517 | Like 1,359 | Not-Interested 1,200 | Combined 1,220 | Total 5,296.

Embeddings

Titles were embedded using Sentence-BERT (all-MiniLM-L6-v2; Reimers and Gurevych 2019). Embeddings were reduced from 768 to 100 dimensions via PCA prior to clustering and t-SNE visualisation.

File Description

allDfEmbed.pkl — Pandas DataFrame (pickle) containing recommendation titles, persona and account labels, SBERT embeddings.

To load:

import pandas as pd
df = pd.read_pickle("allDfEmbed.pkl")

Ethics

This study received institutional ethics approval from the University of Oxford (CUREC Ref: 1562454). Data collection used only newly created accounts with age-appropriate content, limiting the risk of affecting real users' recommendations. The study followed Oxford CUREC Best Practice Guidance 06 for internet-mediated research.

Citation

If you use this dataset, please cite the accompanying paper:

@article{cho2026youtube,
title = {Evaluating feedback mechanisms for aligning {YouTube} recommendations with user interests},
author = {Cho, Desiree and Hale, Scott A. and Zhao, Jun and Shadbolt, Nigel and Przybylski, Andrew K.},
year = {2026},
note = {Manuscript under review}
}

Authors

Desiree Cho, Scott A. Hale, Jun Zhao, Nigel Shadbolt, Andrew K. Przybylski Oxford Internet Institute / Department of Computer Science / Institute for Ethics in AI, University of Oxford Correspondence: desiree.cho@cs.ox.ac.uk