Ranking Model Training via Implicit Feature Extraction

The patent introduces an implicit feature extraction mechanism as an intermediary between source domain data and the ranking model. Instead of directly adding source domain samples to target domain training data, the system extracts implicit features from source domain user behavior data and uses these features to augment the target domain training samples. This intermediary approach allows the model to benefit from additional training signals while avoiding the negative transfer caused by direct sample mixing, as the implicit features capture transferable patterns without introducing distribution inconsistency.

The patent extracts implicit features from source domain data, separating the useful generalizable patterns from the problematic distribution characteristics. By taking out only the essential behavioral patterns encoded in implicit features rather than using complete source domain samples, the method achieves model generalization improvement while avoiding negative transfer from distribution inconsistency between source and target domains.