Neural Text Encoder for Cold Start Page Ranking

The patent introduces content vectors as an intermediary representation that bridges the gap between interaction data and page content. The neural network text encoder transforms page content into vector representations, which then serve as the basis for recommendation when interaction data is insufficient. This mediator enables cold start pages to be recommended without relying solely on interaction data.

The patent changes the parameter used for recommendation from interaction data alone to a combination of content vectors and interaction data. By transforming page content into vector space and comparing it with user interaction patterns, the system can recommend cold start pages that would otherwise be invisible to traditional interaction-based recommendation systems.

The patent performs preliminary encoding of page content into vectors before the recommendation process. This pre-processing step creates a rich representation of page content that can be quickly compared with user interactions, allowing cold start pages to be evaluated and ranked alongside warm pages from the outset, rather than being filtered out due to lack of interaction history.

The patent replaces the mechanical counting of interactions with a neural network-based semantic comparison system. Instead of simply counting how many times a page was clicked, the system uses neural networks to compare the semantic meaning of page content with user interaction patterns, enabling meaningful recommendations for pages with minimal interaction data.