When Bits Break Recourse: Counterfactual-Faithful Quantization
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910228314849280 |
|---|---|
| author | Yahyati, Chaymae Lamaakal, Ismail Makkaoui, Khalid El Ouahbi, Ibrahim |
| author_facet | Yahyati, Chaymae Lamaakal, Ismail Makkaoui, Khalid El Ouahbi, Ibrahim |
| contents | Quantization can preserve predictive accuracy under low-bit deployment while silently breaking algorithmic recourse: an actionable change that flips a decision before quantization may fail after quantization, or become substantially more costly. We formalize counterfactual sensitivity under quantization through validity, cost, and direction stability, and introduce two metrics: Validity Drop (VD) and Counterfactual Recourse Gap (CRG) that reveal recourse failures invisible to accuracy. We propose Counterfactual-Faithful Quantization (CFQ), which trains quantizer parameters and mixed-precision bit allocation to preserve counterfactual behavior by enforcing the target outcome at teacher recourse points under a global bit budget. A margin-based analysis gives a sufficient condition for recourse transfer under bounded quantization perturbations. Experiments on Adult, German Credit, and COMPAS show that accuracy-matched baselines can significantly degrade recourse stability, while CFQ maintains accuracy and substantially improves VD and CRG across bit budgets. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_17160 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | When Bits Break Recourse: Counterfactual-Faithful Quantization Yahyati, Chaymae Lamaakal, Ismail Makkaoui, Khalid El Ouahbi, Ibrahim Machine Learning Artificial Intelligence Computer Vision and Pattern Recognition Quantization can preserve predictive accuracy under low-bit deployment while silently breaking algorithmic recourse: an actionable change that flips a decision before quantization may fail after quantization, or become substantially more costly. We formalize counterfactual sensitivity under quantization through validity, cost, and direction stability, and introduce two metrics: Validity Drop (VD) and Counterfactual Recourse Gap (CRG) that reveal recourse failures invisible to accuracy. We propose Counterfactual-Faithful Quantization (CFQ), which trains quantizer parameters and mixed-precision bit allocation to preserve counterfactual behavior by enforcing the target outcome at teacher recourse points under a global bit budget. A margin-based analysis gives a sufficient condition for recourse transfer under bounded quantization perturbations. Experiments on Adult, German Credit, and COMPAS show that accuracy-matched baselines can significantly degrade recourse stability, while CFQ maintains accuracy and substantially improves VD and CRG across bit budgets. |
| title | When Bits Break Recourse: Counterfactual-Faithful Quantization |
| topic | Machine Learning Artificial Intelligence Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2605.17160 |