Recommender System Guidelines

Introduction

Ollage uses recommender systems to help you discover clips and creators that match your interests. This page explains how our recommender works, what factors influence what you see, and what choices you have as a user. This information is provided in accordance with the EU Digital Services Act (DSA) and applicable transparency regulations.

What Recommendation Features We Offer

Personalized "For You" feed: If you are logged in and have interacted with clips (liked, saved, excluded, or grouped), Ollage shows you a personalized ranking of clips based on your taste profile. This is the default experience.

Anonymous browsing: If you browse without logging in (or reset your taste data), you will see a non-personalized ranking based on platform-wide signals. You may also switch to this mode at any time.

Browse by creator or tag: You can filter and sort clips by creator, tag, or other structured dimensions using the universal filter bar.

Main Ranking Signals

Ollage ranks clips using four layers of signals:

Layer 1: Discovery OPS (Platform Signals)

Our baseline ranking uses Naive Bayes tag log-likelihood ratios derived from RedGifs platform popularity data (configuration file: t1_tag_weights.json). This layer scores clips based on how strongly their tags correlate with community engagement across the platform. It captures broad trends and is applied to all clips.

Layer 2: Content-Value Model (CVM)

A logistic-regression model (cvm_weights.json) combines Discovery OPS scores with clip-level signals including:

• View count
• Like count
• Percentage watched (how far viewers watch before stopping)
• Recency (freshness of content)
• Duration

The CVM estimates which clips are most likely to be curated as high-quality. It applies globally to all users.

Layer 3: Creator Niche Affinity

A TF-IDF similarity model (creator_model.json) compares the tag profile of a clip's creator against the tag profiles of creators in Ollage's curated network. This signal is used for fast triage and prioritization, not automatic acceptance or rejection. It does not personalize recommendations.

Layer 4: CLIP Vision Probe (Per-User Personalization)

A logistic-regression model (clip_probe.json, AUC 0.73) trained on clip thumbnail embeddings (OpenAI ViT-B-32) learned from human comparison votes. When you are logged in, this layer personalizes clip scores based on:

• Your taste history: clips you've liked, saved, excluded, or grouped over time
• Your interaction patterns: how far you watch clips, how often you interact with specific creators or tags
• Your stated preferences: groups (albums) you create and organize manually

The vision probe learns from your behavior and retrains every 100 human votes you cast across the platform (votes in tournaments, comparisons, or direct film-feedback). The model blends a global "baseline taste" (learned from all users) with your personal centroid (learned from your interactions) using a clone-then-fork approach. This ensures new users see broad popular content while experienced users see increasingly personalized recommendations.

Your Controls Over Recommendations

You directly influence your recommendations by:

• Liking and saving clips — these actions update your taste profile immediately
• Excluding clips — marking clips as "not for me" trains the model to deprioritize similar content
• Creating and organizing groups (albums) — sorting clips into numbered albums teaches the model your aesthetic preferences
• Voting in tournaments and comparisons — head-to-head votes retrain the vision probe and directly affect your "For You" ranking
• Watching time — how far you watch a clip (if you finish it, skip to the end, or close it) signals to the model whether it matched your interests

Opting Out of or Resetting Personalization

Clear your taste data: In your account settings, you can reset all your taste history (likes, saves, exclusions, groups, votes). This immediately reverts you to the non-personalized ranking for all future sessions.

Browse anonymously: Log out or use your browser's private/incognito mode. Anonymous sessions use the bb_session cookie to track votes for deduplication but do not merge with any account. You can reset this cookie or clear browser data between sessions.

Non-personalized alternative: Ollage always offers a global "trending" ranking available without logging in. This ranking uses only Discovery OPS, CVM, and Creator Niche Affinity signals — no personalization.

What We Suppress from Recommendations

The following content does not appear in any recommendation:

• Content violating Ollage's policies: All content is screened against the Acceptable Use Policy. Prohibited content is removed and never recommended.
• Low-quality clips: Clips scoring below the CVM quality threshold are deprioritized or excluded from recommendations (media-quality gate: bitrate, resolution, duration, frame rate, and histogram analysis).
• Reported or flagged content: Clips you have reported or that are under review for policy violations do not appear in recommendations until review is complete.
• DMCA or takedown content: Clips removed due to copyright notices or legal takedown are excluded permanently.

Transparency and No Automated Legal Decisions

No fully automated legal or similarly significant decisions: Ollage does not use the recommender system to make decisions with legal effect or similarly significant consequences (such as account termination, payment processing, or credit decisions). All account moderation decisions are reviewed by humans.

Request explanations: If you believe the recommendation system has incorrectly ranked or excluded a clip, you can contact us at [email protected] and request an explanation. We will respond within 7 days.

Data Retention and Your Taste History

Your taste history (likes, saves, exclusions, groups, votes) is stored in the taste_events table and is retained as long as your account exists. Anonymous browsing cookies (bb_session) are retained for 90 days.

For more information about how we use your data, see our Privacy Policy. To delete your taste history without deleting your account, visit your account settings.

Algorithm Updates and Changes

Ollage may update the recommender system from time to time to improve quality, address bias, or comply with legal requirements. Significant changes to the algorithm or signals will be communicated via in-app notifications or email. Your taste history will not be erased due to algorithm updates unless you explicitly request it.

Questions or Concerns

If you have questions about how the recommender system works, or if you wish to:

• Request your taste history data: Email [email protected]
• Request explanations for recommendations or exclusions: Email [email protected]
• Report abuse or misuse of the recommender system: Email [email protected] or use the in-app report button on any clip
• Submit feedback on recommendations quality: Use the feedback form in your account or email [email protected]