Search Relevance Scoring via Vector Embeddings and User Behavior

The patent transforms search term representations from simple frequency counts to vector embeddings in a multi-dimensional space. Each search term is converted to a vector that captures contextual relationships, allowing the system to measure similarity through vector distance metrics rather than simple term frequency matching. This parameter transformation enables the system to model user-perceived relevance more accurately while maintaining computational efficiency.

The patent introduces vector embeddings as an intermediary representation layer between raw search terms and ranking decisions. Instead of directly comparing term frequencies, the system first converts terms to vector representations that encode contextual meanings, then uses these vectors to compute similarity scores. This intermediary transformation allows the system to capture nuanced relationships between search terms and results that traditional TF-IDF cannot detect.

The patent combines multiple user behavior signals (click frequency, bid frequency, skip frequency, dwell time) into a unified training dataset for learning models. Instead of maintaining separate complex data structures for each signal type, the system integrates them into a consolidated format that feeds into vector embedding generation and similarity computation, reducing overall system complexity while preserving the informational value of each signal.

The system automatically generates vector embeddings and computes similarity scores using learning models trained on user behavior data, without requiring manual configuration or complex rule-based processing. The learning models self-adapt to user preferences through continuous training on behavioral signals, reducing the need for manual data structure management and simplifying the overall system architecture.

The patent pre-trains vector embedding models and learns similarity relationships between search terms and potential results before actual queries are processed. By performing the computationally intensive model training and vector space construction in advance, the system reduces the processing time required for individual queries, as the heavy computational work has already been completed during the pre-training phase.

The system dynamically adjusts the level of computational processing based on query characteristics and available resources. For simple queries, the system may use pre-computed vector similarities, while for more complex queries, it can invoke full learning model inference. This dynamic approach allows the system to balance relevance accuracy with processing time requirements, adapting its computational effort to the specific needs of each query.