The Phenomenal World

November 18th, 2017

The Phenomenal World

Duchamp Wanted



How to build justice into algorithmic actuarial tools

Key notions of fairness contradict each other—something of an Arrow’s Theorem for criminal justice applications of machine learning.

"Recent discussion in the public sphere about algorithmic classification has involved tension between competing notions of what it means for a probabilistic classification to be fair to different groups. We formalize three fairness conditions that lie at the heart of these debates, and we prove that except in highly constrained special cases, there is no method that can satisfy these three conditions simultaneously. Moreover, even satisfying all three conditions approximately requires that the data lie in an approximate version of one of the constrained special cases identified by our theorem. These results suggest some of the ways in which key notions of fairness are incompatible with each other, and hence provide a framework for thinking about the trade-offs between them."

Full paper from JON KLEINBERG, SENDHIL MULLAINATHAN and MANISH RAGHAVAN here. h/t research fellow Sara, who recently presented on bias in humans, courts, and machine learning algorithms, and who was the source for all the papers in this section.

In a Twitter thread, ARVIND NARAYANAN describes the issue in more casual terms.

"Today in Fairness in Machine Learning class: a comparison of 21 (!) definitions of bias and fairness [...] In CS we're used to the idea that to make progress on a research problem as a community, we should first all agree on a definition. So 21 definitions feels like a sign of failure. Perhaps most of them are trivial variants? Surely there/s one that's 'better' than the rest? The answer is no! Each defn (stat. parity, FPR balance, contextual fairness in RL...) captures something about our fairness intuitions."

Link to Narayanan’s thread.

Jay comments: Kleinberg et al. describe their result as choosing between conceptions of fairness. It’s not obvious, though, that this is the correct description. The criteria (calibration and balance) discussed aren’t really conceptions of fairness; rather, they’re (putative) tests of fairness. Particular questions about these tests aside, we might have a broader worry: if fairness is not an extensional property that depends upon, and only upon, the eventual judgments rendered by a predictive process, exclusive of the procedures that led to those judgments, then no extensional test will capture fairness, even if this notion is entirely unambiguous and determinate. It’s worth consideringNozick’s objection to “pattern theories” of justice for comparison, and (procedural) due process requirements in US law.

⤷ Full Article

November 11th, 2017

The Hülsenbeck Children



Recommender systems power YouTube's controversial kids' videos

Familiar cartoon characters are placed in bizarre scenarios, sometimes by human content creators, sometimes by automated systems, for the purpose of attracting views and ad money. First, from the New York Times:

“But the app [YouTube Kids] contains dark corners, too, as videos that are disturbing for children slip past its filters, either by mistake or because bad actors have found ways to fool the YouTube Kids algorithms.

“In recent months, parents like Ms. Burns have complained that their children have been shown videos with well-known characters in violent or lewd situations and other clips with disturbing imagery, sometimes set to nursery rhymes. Many have taken to Facebook to warn others, and share video screenshots showing moments ranging from a Claymation Spider-Man urinating on Elsa of ‘Frozen’ to Nick Jr. characters in a strip club.”

Full piece by SAPNA MAHESHWARI in the Times here.

On Medium, JAMES BRIDLE expands on the topic, and criticizes the structure of YouTube itself for incentivizing these kinds of videos, many of which have millions of views.

“These videos, wherever they are made, however they come to be made, and whatever their conscious intention (i.e. to accumulate ad revenue) are feeding upon a system which was consciously intended to show videos to children for profit. The unconsciously-generated, emergent outcomes of that are all over the place.

“While it is tempting to dismiss the wilder examples as trolling, of which a significant number certainly are, that fails to account for the sheer volume of content weighted in a particularly grotesque direction. It presents many and complexly entangled dangers, including that, just as with the increasing focus on alleged Russian interference in social media, such events will be used as justification for increased control over the internet, increasing censorship, and so on.”

Link to Bridle’s piece here.

⤷ Full Article

November 4th, 2017




Sociologist Zeynep Tufekci engages with Adam Mosseri, who runs the Facebook News Feed

Tufekci: “…Facebook does not ask people what they want, in the moment or any other way. It sets up structures, incentives, metrics & runs with it.”

Mosseri: “We actually ask 10s of thousands of people a day how much they want to see specific stories in the News Feed, in addition to other things.”

Tufekci: “That’s not asking your users, that’s research on your product. Imagine a Facebook whose customers are users—you’d do so much differently. I mean asking all people, in deliberate fashion, with sensible defaults—there are always defaults—even giving them choices they can change…Think of the targeting offered to advertisers—with support to make them more effective—and flip the possibilities, with users as customers. The users are offered very little in comparison. The metrics are mostly momentary and implicit. That’s a recipe to play to impulse.”

The tweets are originally from Zeynep Tufekci in response to Benedict Evans (link), but the conversation is much easier to read in Hamza Shaban’s screenshots here.

See the end of this newsletter for an extended comment from Jay.

  • On looping effects (paywall): “This chapter argues that today's understanding of causal processes in human affairs relies crucially on concepts of ‘human kinds’ which are a product of the modern social sciences, with their concern for classification, quantification, and intervention. Child abuse, homosexuality, teenage pregnancy, and multiple personality are examples of such recently established human kinds. What distinguishes human kinds from ‘natural kinds’, is that they have specific ‘looping effects’. By coming into existence through social scientists' classifications, human kinds change the people thus classified.” Link. ht Jay


Mechanisms and causes between micro and macro

Daniel Little, the philosopher of social science behind Understanding Society, haswritten numerous posts on the topic. Begin with this one from 2014:

“It is fairly well accepted that there are social mechanisms underlying various patterns of the social world — free-rider problems, communications networks, etc. But the examples that come readily to mind are generally specified at the level of individuals. The new institutionalists, for example, describe numerous social mechanisms that explain social outcomes; but these mechanisms typically have to do with the actions that purposive individuals take within a given set of rules and incentives.

“The question here is whether we can also make sense of the notion of a mechanism that takes place at the social level. Are there meso-level social mechanisms? (As always, it is acknowledged that social stuff depends on the actions of the actors.)”

In the post, Little defines a causal mechanism and a meso-level mechanism, then offers example research.

“…It is possible to identify a raft of social explanations in sociology that represent causal assertions of social mechanisms linking one meso-level condition to another. Here are a few examples:

  • Al Young: decreasing social isolation causes rising inter-group hostility (link)
  • Michael Mann: the presence of paramilitary organizations makes fascist mobilization more likely (link)
  • Robert Sampson: features of neighborhoods influence crime rates (link)
  • Chuck Tilly: the availability of trust networks makes political mobilization more likely (link)
  • Robert Brenner: the divided sovereignty system of French feudalism impeded agricultural modernization (link)
  • Charles Perrow: legislative control of regulatory agencies causes poor enforcement performance (link)

More of Little’s posts on the topic are here. ht Steve Randy Waldman

⤷ Full Article