Semantic Relevance Inference for Absolute Search Ranking

The patent transforms the relevance scoring parameter from relative scores (context-dependent, requiring conversion) to absolute scores (independent, directly usable). This is achieved by training machine learning models to output absolute relevance scores that represent the true relevance of content items to search queries, eliminating the need for complex conversion processes and improving measurement precision.

The patent replaces the mechanical scoring system (rule-based relative scoring) with an intelligent system (machine learning models) that automatically learn and generate absolute relevance scores. This substitution enables the system to achieve higher accuracy in relevance determination while the models handle the complexity internally, providing clean absolute scores to the retrieval system.

The patent implements a feedback mechanism where user interactions (clicks, dwell time, selections) are collected and used to retrain the machine learning models. This continuous feedback loop allows the system to learn from actual user behavior and improve relevance ranking accuracy over time. The feedback is processed efficiently by updating model parameters incrementally, balancing accuracy improvement with acceptable retraining time.

The patent performs preliminary actions by pre-processing user feedback data and preparing training datasets before model retraining. This includes aggregating user interactions, filtering relevant feedback, and formatting data in advance, which reduces the actual retraining time and computational burden when model updates are needed.