IEEE Technology and Society Magazine - September 2021 - 81

formation applied to the ML pipeline. Problem formulation
[6] can be understood as an iterative and
dynamic process applied to specific data science
projects. That is, problem formulation is not static
but rather requires constant reformulation and normative
reevaluation of the relation between questions,
choices, decisions, and goals by key actors.
In this respect, the activity of problem formulation
has an irreducible epistemological and ethical reality
which cannot be dissolved by a strictly solutionist
approach. A solutionist approach treats every problem
instrumentally as an unreflective given and an
unfortunate accident that should be eliminated by a
good and necessary solution. Problem formulation
counters the solutionist approach by acknowledging
that ML problems exist in contexts of epistemic
ambiguity. That is, before identifying a solution, one
needs to carefully understand how a problem is constituted
and formulated3 and whether the proposed
solution is sufficient given the epistemological and
social context that one operates in. The question
here is whether the problem is robustly constituted
or whether it lacks sufficient conditions that are
attuned to a given context. In this respect, the problem
is not merely given but it is made based on the
conditions of epistemic ambiguity.
To further illustrate this point, we will focus now
on an example of problem formulation on the LinkedIn
platform. Our example is based on an interview
with Romer Rosales on the topic of problem formulation
and the selection of the right objective function
for a given problem. Rosales and Charrington
[7] develop an approach that consists of algorithmic
modeling of " holistic metrics. " Their stated objective
is to increase user engagement on the platform. The
means to do so is via high-frequency emails sent to
LinkedIn members. This approach turns out to be
misguided as measured by the negative feedback in
response to the over-reliance on email to increase
user participation. In this case, negative feedback
can be conceptualized in terms of a computational
practice of quantification of reactional costs
on the platform which functions as an indicator to
measure the levels of satisfaction of the platforms'
users. Understanding the negative feedback in terms
of reactional costs has the advantage to place the
3 In a recent online talk given in December 2020 at the Montreal.AI Debate 2,
Dr. Fei-Fei Li described her account of problem formulation as seeking a North
Star in the following terms: " The mere formulation of a problem is often far more
essential than its solution, which [...] requires creative imagination and marks real
advances in science. " https://www.youtube.com/watch?v=XY1VTLRIsNo&t=59s
users' reaction in an internal schema of contestability
which allows the LinkedIn practitioners to reformulate
the problem. The reason for addressing what
is in retrospect a poor problem formulation is due to
the exclusive focus of achieving a narrow short-term
goal of increasing user participation by focusing
exclusively on maximizing one particular metric at
the expense of others.
Instead, the new problem reformulation is articulated
in terms of optimizing competing metrics. The
LinkedIn team thus decides to reduce the maximization
of email-based opportunities for connecting
on the platform and instead focuses on emphasizing
" valuable engagement " or by facilitating the experience
of a meaningful connection from the perspective
of the user. Here, we have two different ways of
approaching user engagement with the platform: a
strictly quantitative one from the perspective of the
network which is predicated on increasing platform
dependence and potential user dissatisfaction and a
" meaningful " one centered on the actual user experience.
Also, at this level, one can note that this new
problem reformulation engendered a significant
temporal shift. The new approach, which facilitates
meaningful user engagement, moved from acquiring
narrow short-term gains (via a mechanism of
evaluation of click-through rate) to providing " longterm
value to the member " that is by increasing
participation on the network in a longer term time
frame. By rejecting the optimization of a single metric
approach, the solution envisioned by the LinkedIn
team is to computationally design strategies to
develop meaningfulness and practices of inclusion
on the platform.
At this point, we can state that problem formulation
works as a practice of redefinition of concepts
or as a practice of rationalization of the unthought
ML practices and discourses. Also, we have here a
practice of redefinition of strategies at an ethical and
epistemological level. In both cases, problem formulation
works as a practice of the computational
modes of knowing which shape our ethical commitments.
What is at stake here is precisely that computational
analysis brings to light questions that were
unthinkable before its emergence as a mode of analysis
and practice. This new mode of knowing not
only challenges our existing ethical frameworks and
underlying assumptions, but also requires that we
create new ethical modes of analysis and practice
September 2021
81
http://www.Montreal.AI https://www.youtube.com/watch?v=XY1VTLRIsNo&t=59s
IEEE Technology and Society Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2021
