Abstract
Generating recommendation signals for media content items that have good coverage and that include details about the media content item may be challenging, especially for a large media content universe. In some cases, sparse signals may fail to capture nuanced relationships between different media content items, leading to less effective media content recommendations. Existing methods for generating more abstract signals may involve significant manual effort and/or complex data extraction processes.
The disclosed technology for an improved recommendation system addresses these limitations by using hierarchical clustering of media content items (e.g., movies, television shows, short form videos, etc.). The recommendation system may generate a hierarchical tree structure by performing hierarchical clustering of media content items. Each node of the tree structure may represent a specific cluster of media content items with higher-level nodes representing increasingly abstract thematic clusters of media content items. Each node may be associated with a score or value representative of the affinity of the user with the corresponding cluster of media content items for a particular topic represented by the node. For example, when a user interacts with and/or selects a media content item, the recommendation system may traverse a path from the leaf representing the media content item to the node that represents the cluster that includes the media content item, and then through one or more nodes to the root of the tree structure. The recommendation system may update the values associated with each node along the traversed path based on the user selection of the media content item. The recommendation system may calculate the affinity of the user for a new media content item by summing the values of the nodes along a path from a leaf representing the new media content item to the node that represents the cluster that includes the new media content item, and then through one or more nodes to the root of the tree structure.
The disclosed technology for the improved recommendation system may create dynamic and detailed hierarchical path-based signal generation for media content recommendations by providing signals that capture both specific media content item properties along with broader thematic connections for the media content item. The resulting recommendation signals may be less sparse and may be adjusted based on the size and nature of the content universe, improving the relevance of media content recommendations.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Chatterjee, Tamojit and Nayak, Shravan, "Hierarchical Path-Based Signal Generation for Content Recommendations", Technical Disclosure Commons, (September 04, 2025)
https://www.tdcommons.org/dpubs_series/8553