Dynamic Content Selection Using Correlation Scores

The system pre-computes and stores embedding vectors for content pieces during an offline training phase using historical interaction data. These pre-computed embeddings are stored in a database, allowing the system to quickly retrieve and compare them with visitor behavior embeddings in real-time without performing complex computations during the online serving phase.

The patent replaces traditional mechanical content selection methods (rule-based filtering, manual categorization) with an unsupervised learning approach using embedding vectors. The system uses cosine similarity computation between visitor behavior embeddings and content embeddings to automatically rank and select relevant content, substituting manual or rule-based mechanisms with automated machine learning-based selection.

The system performs offline training to learn embedding vectors from available historical data, creating a pre-trained model that can generalize to new visitors. Even when individual visitor history is limited, the pre-trained embeddings capture general patterns of visitor behavior and content preferences, enabling reliable content recommendation without requiring extensive individual historical data.

The patent uses embedding vectors as compressed representations (copies) of complex visitor behavior patterns and content characteristics. Instead of storing and processing raw historical interaction data for each visitor, the system creates compact embedding vectors that capture essential features, allowing efficient comparison and matching while reducing data storage and processing requirements.