Lydia T. Liu, Sarah Dean, Esther Rolf, Max Simchowitz, Moritz Hardt
2018 · arXiv
Fairness in machine learning has predominantly been studied in static classification settings without concern for how decisions change the underlying population over time. Conventional wisdom suggests that fairness criteria promote the long-term well-being of those groups they aim to protect. We study how static fairness criteria interact with temporal indicators of well-being, such as long-term improvement, stagnation, and decline in a variable of interest. We demonstrate that even in a one-step feedback model, common fairness criteria in general do not promote improvement over time, and may in fact cause harm in cases where an unconstrained objective would not. We completely characterize the delayed impact of three standard criteria, contrasting the regimes in which these exhibit qualitatively different behavior. In addition, we find that a natural form of measurement error broadens the regime in which fairness criteria perform favorably. Our results highlight the importance of measurement and temporal modeling in the evaluation of fairness criteria, suggesting a range of new challenges and trade-offs.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
3 selected excerpts · 3 total mentions
Reducing accuracy to try and improve fairness is definitely the politically correct thing to do. There was a good paper on this topic about how trying to make a model more fair can result in worst outcomes: https://arxiv.org/abs/1803.04383 If some group does happen to have a higher recidivism risk, pretending this isn't true and just correcting against that fact will just result in more crime.
Fairness, Accountability, Transparency (aka FAT) is a real sub-field of machine learning, that is as technical as any other machine learning sub-field in my opinion. It is not a PR stunt. I've published at top ML conference a paper about some CUDA kernels I wrote to accelerate training a class of RNNs, so I feel have technical grounding to make these claims. Here are some FAT papers I've enjoyed: https://arxiv.org/abs/1806.08010 https://arxiv.org/abs/1802.04023 https://arxiv.org/abs/1803.04383 (won best paper award at ICML 2018) Many FAT papers are published at NeurIPS or ICML (generally considered top two machine learning conferences). There's also a conference just on the topic: https://facctconference.org/
> The problem is the self-reinforcing, external factors that these algorithms don't take into account. And changing the algorithms isn't the right way to fix those external factors. If you try, their self-reinforcing nature is going to cause exactly the opposite of what you want. To stay with the crime vs. race example, assume that the ratio of black to white criminals is greater than you'd expect from the ratio of blacks to whites in the overall population, and that is due to self-reinforcing external factors. Your algorithm for deciding whether to parole a prisoner picks up on that statistical regularity and rates blacks riskier than whites. Because you want to achieve…
representational fairness, you make the algorithm race-blind to equalize the rates. What happens? The black prison population shrinks. But your algorithm wasn't wrong when it predicted higher recidivism risk, so the black crime rate rises. Your algorithm now requires an even stronger correction to remain race-blind. Black criminals realize that they have an easier time getting paroled, so the black crime rate rises some more. Debiasing the algorithm has made the situation worse. In general, when you find out that your machine-learning model has some undesirable bias, the correct response is not to blind the algorithm. Instead, you should fix the data-generation process (the real world) until your model no longer picks up on the bias. Edit: For a mathematically precise description of how attempts to make an algorithm more fair cause worse results when you consider the feedback loop that it is embedded in, see "Delayed Impact of Fair Machine Learning" https://arxiv.org/abs/1803.04383
More from the same topics.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Chiyuan Zhang, Samy Bengio, Moritz Hardt +2
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
A. Pluchino. A. E. Biondo, A. Rapisarda
The largely dominant meritocratic paradigm of highly competitive Western cultures is rooted on the belief that success is due mainly, if not exclusively, to personal qualities such as talent, intelligence, skills, efforts or risk taking.