How a Recommendation System Learns What You Really Want to Watch

This patent describes a media recommendation system that creates a unique user profile, uses it to pick a starting point, and then dynamically adjusts a list of suggestions based on what you've already watched and your real-time feedback.

What it covers

The system first builds a "profile vector" for a user (Claim 1). This profile is clever because it gives less weight to media items the user has watched many times, encouraging new discoveries. For example, if you've watched "The Matrix" 50 times, that movie's influence on your profile might be reduced. Next, it picks a "seed media item" based on this profile, which acts as a starting point for recommendations (Claim 1). This seed could be a trending show or something you recently liked (Claim 4). Then, the system creates a "pool" of other media items and sorts them by how similar they are to the seed item (Claim 1). Finally, it selects an item from this sorted pool for you to watch (Claim 1). Crucially, if you give feedback, like skipping a song or rating a movie, the system instantly updates the pool with new suggestions (Claim 3).

What it doesn't cover

• —Does not cover recommendation systems that do not reduce the influence of frequently consumed media items in the user profile.
• —Does not cover systems that generate recommendations without first establishing a "seed media item" based on the user's profile.
• —Does not cover systems where the pool of media items is not sorted by vector distance to a seed item.
• —Does not cover systems that do not dynamically modify the recommendation pool based on user feedback.
• —Does not cover systems that rely solely on explicit user ratings without considering consumption frequency.

The clever bit

The truly novel aspect is how the user's profile is generated: it's "inversely proportional to a total number of times that the user has consumed a specific media item" (Claim 1). This means the system actively reduces the weight of content you've watched repeatedly, pushing it to recommend fresh, relevant items rather than endlessly suggesting your all-time favorites.

Why it matters

Recommendation systems are fundamental to how users discover content across streaming platforms, social media, and e-commerce. This patent aims to refine how these systems learn user preferences, potentially leading to more engaging and less repetitive content discovery experiences. By dynamically adapting to user feedback and de-emphasizing over-consumed content, it seeks to improve user satisfaction and retention in competitive media markets.

Real-world examples

• 1.Netflix's "Because you watched..." suggestions
• 2.Spotify's Discover Weekly playlist
• 3.YouTube's video recommendations
• 4.TikTok's For You Page algorithm
• 5.Amazon's "Customers who bought this also bought..."