🤖 AI Summary
Neural information retrieval (IR) faces significant challenges in machine unlearning—specifically, non-normalized relevance scores hinder teacher-student distillation, while query/document entanglement between forget and retain sets degrades ranking performance. Method: This work formally introduces the novel task of Neural Machine Unranking (NuMuR) and proposes Contrastive Consistency Loss (CoCoL), which jointly integrates contrastive learning and consistency regularization. To enable efficient, targeted unlearning, CoCoL is coupled with gradient masking and lightweight fine-tuning. Contribution/Results: The approach ensures strong controllability of forgetting while preserving ranking consistency. Evaluated on multiple standard IR benchmarks, it substantially outperforms state-of-the-art indiscriminate unlearning methods, achieving new SOTA performance in both targeted unlearning accuracy and retention of downstream ranking effectiveness.
📝 Abstract
We tackle the problem of machine unlearning within neural information retrieval, termed Neural Machine UnRanking (NuMuR) for short. Many of the mainstream task- or model-agnostic approaches for machine unlearning were designed for classification tasks. First, we demonstrate that these methods perform poorly on NuMuR tasks due to the unique challenges posed by neural information retrieval. Then, we develop a methodology for NuMuR named Contrastive and Consistent Loss (CoCoL), which effectively balances the objectives of data forgetting and model performance retention. Experimental results demonstrate that CoCoL facilitates more effective and controllable data removal than existing techniques.