Fortress Framework Stabilizes Search Recommendations via Feature Pruning

A novel framework named Fortress is designed to enhance the reliability and precision of predictive models utilized in search and recommendation systems. It tackles the issue of temporal instability, which arises from input features that can cause fluctuations in output scores, ultimately affecting model dependability and user satisfaction, particularly in multi-stage systems. Fortress employs historical snapshots—datasets divided by time that reflect score variations for the same entity over different periods—to pinpoint and eliminate features that lead to erratic predictions. This method consists of four phases: gathering historical snapshots, detecting samples with unstable predictions, isolating and discarding features that induce instability, and retraining models with only stable features. The approach incorporates semantic features from LLMs and BERT-based models to enhance generalization. The paper can be found on arXiv with the identifier 2605.15299.