Scholarlink

<p>Machine learning is everywhere now-lending, criminal justice, healthcare-and with that comes a real problem: transparency and fairness. Deep learning models can be powerful, but they are also black boxes. People affected by these decisions have no clear understanding of how they are being judged. Organizations cannot catch bias. Regulators cannot tell if anyone is following the rules. There is this idea floating around that one has to pick between accuracy and interpretability-you cannot have both. This work challenges that assumption directly. That is where NAM++ comes in. This model hits three big marks: it is just as accurate as deep learning models people already use, it actually shows how each feature affects decisions, and it keeps things fair across different measures. The magic is in how it builds on additive models but throws in the right feature interactions, so one gets the complex stuff without losing the ability to explain what is happening. To back up the accuracy, bootstrap methods were used, not just single number comparisons, so the results hold up. For fairness, several definitions were looked at, since not everyone agrees on what fair really means. NAM++ was tried on two real world datasets that matter for regulations. The Adult Income dataset has 45,222 cases. NAM++ scored an AUC of 0.9106, pretty much tied with the standard deep network’s 0.9093, with a confidence interval that includes zero. That means there is no real difference. On the ProPublica COMPAS dataset (6,857 cases), NAM++ got 0.7245 AUC, while other models landed at 0.7210, again statistically tied. And across both datasets, NAM++ held its ground or even improved when it came to fairness-metrics like Demographic Parity and Equalized Odds. Training did not need anything fancy, either-ten to twelve cycles, regular hardware, and could be performed without difficulty. So the idea that one has to trade interpretability for accuracy just does not hold up here. By making transparency and fairness core goals from the start, not tacked on at the end, this work shows one can build powerful, explainable, and fair systems. That is a big deal, especially in areas where trust and meeting the rules are not optional.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Nam++: achieving interpretability and fairness without sacrificing accuracy through neural additive models with selective feature interactions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月25日

查看全文
  • A. Sushiiel

摘要

Machine learning is everywhere now-lending, criminal justice, healthcare-and with that comes a real problem: transparency and fairness. Deep learning models can be powerful, but they are also black boxes. People affected by these decisions have no clear understanding of how they are being judged. Organizations cannot catch bias. Regulators cannot tell if anyone is following the rules. There is this idea floating around that one has to pick between accuracy and interpretability-you cannot have both. This work challenges that assumption directly. That is where NAM++ comes in. This model hits three big marks: it is just as accurate as deep learning models people already use, it actually shows how each feature affects decisions, and it keeps things fair across different measures. The magic is in how it builds on additive models but throws in the right feature interactions, so one gets the complex stuff without losing the ability to explain what is happening. To back up the accuracy, bootstrap methods were used, not just single number comparisons, so the results hold up. For fairness, several definitions were looked at, since not everyone agrees on what fair really means. NAM++ was tried on two real world datasets that matter for regulations. The Adult Income dataset has 45,222 cases. NAM++ scored an AUC of 0.9106, pretty much tied with the standard deep network’s 0.9093, with a confidence interval that includes zero. That means there is no real difference. On the ProPublica COMPAS dataset (6,857 cases), NAM++ got 0.7245 AUC, while other models landed at 0.7210, again statistically tied. And across both datasets, NAM++ held its ground or even improved when it came to fairness-metrics like Demographic Parity and Equalized Odds. Training did not need anything fancy, either-ten to twelve cycles, regular hardware, and could be performed without difficulty. So the idea that one has to trade interpretability for accuracy just does not hold up here. By making transparency and fairness core goals from the start, not tacked on at the end, this work shows one can build powerful, explainable, and fair systems. That is a big deal, especially in areas where trust and meeting the rules are not optional.