Conference Program
Program - Day 1
8.45 - Opening
Session 1.1: Meaningfulness, usefulness and legitimacy: definitions and concepts
The talk introduces a framework expanding the notion of meaningfulness introduced in measurement theory. We introduce and explain the notions of usefulness and legitimacy of information. Through a number of examples we show that these three requirements are logically independent, this implying that all three of them need to be "checked" before using any (formally established) information within a decision process.
Index numbers are widely used to understand trends, make comparisons, and support decisions. Some statements using index numbers are meaningful but useless, others are meaningful but not legitimate, and so on. We will explore these ideas using two examples, body mass indices and air pollution indices. We will explore the meaningfulness, usefulness, and legitimacy of different kinds of conclusions using such indices. We will also explore conclusions as to whether a more useful or legitimate index can replace a less useful or legitimate one, which includes analysis of the meaningfulness, usefulness, and legitimacy of statements involving correlations and regression relations between indices.
10.00 - Meaningless but useful information
▼
The meaningfulness of a statement cannot be assessed without some information about the measurement scale involved in the statement. This information is usually a uniqueness statement about the measurement scale. In this paper, I claim that statements supplemented with a uniqueness statement can be meaningless but useful if they are supplemented with more information.
10.30 - Coffee
Session 1.2: Argumentation
Deliberated judgments have been introduced by Meinard and myself as representing the judgments of a subject towards a topic; judgments that she reaches after having considered relevant arguments about that topic. In this talk I consider deliberated preferences, which can be viewed as judgments on binary comparison issues. It is meant to contrast to the usual notion of (shallow) preferences that reflect the everyday preferences of individuals. I will focus on one aspect of theories of deliberated preferences, namely, their falsifiability. I will describe what is meant by falsifiability in this context and argue for the need of falsifiable theories. Following the axiomatic method, I will present multiple approaches to define this notion formally and deduce properties that some of these definitions permit to satisfy. I will argue that these properties help define what forms theories of deliberated preferences should adopt.
The paper discusses the collision of different argumentation frames, specifically between mainstream media and the judiciary world, as well as scientific disputes. It highlights that different groups have different criteria for accepting or rejecting arguments, and that the mediatic world tends to reduce the length of argumentative discussion. We focus on a case study, the Liebeck v. McDonald's legal case, to illustrate the impact of media discourse on different argumentation spaces. We'll show that arguments within specific spaces may not make sense to those outside of it due to differences in language, rules of evidence, and reasoning.
Argumentation Frameworks are a useful tool for modelling argumentative dynamics in agent-based models, and the recent introduction of gradual argumentation semantics expands their potential. Our research focuses on using these tools to examine an epistemic community of agents who use the result of their experimentation to produce arguments and exchange them with their peers. The central objectives are to analyze the stability and convergence of opinions among agents and evaluate individual and collective epistemic success.
12.30 - Lunch
Session 1.3: Case study 1
14.30 - Case studies (Minister Ecology)
16.00 - Coffee
Session 1.4: Indexes
Today, most governments are concerned about the well-being or happiness of their citizens, and this is certainly a positive development. In France, the "Stiglitz, Sen, and Fitoussi Report," commissioned in 2009 by Nicolas Sarkozy, emphasized the importance of considering indicators other than GDP to evaluate public policies. In response to this report, the National Institute of Statistics and Economic Studies (INSEE) and the Service for Observation and Statistics announced their intention in 2010 to incorporate dimensions of sustainable development and well-being into economic statistics. This approach is part of a much broader trend and is among other national and international initiatives, including from the United Kingdom, European Union and the United Nations.
The main challenge that economists have to contend with, however, lies in the highly subjective nature of the phenomenon being measured. The vast majority of indicators used are constructed from satisfaction surveys, the biases of which are well-known. However, scientific studies and happiness indicators are proliferating. There is now talk of happiness economics or happiness psychology (or more broadly, positive psychology), or even more generally, the science of happiness. A scientific journal called "The Journal of Happiness Studies" was even created in 2000.
The purpose of this study is to examine the philosophical foundations of these indicators. What concept of well-being or happiness are we talking about? To what extent can well-being truly be assessed? And finally, what is the normative significance of these indicators? A thorough examination of these questions is necessary to avoid falling into the trap of measuring something without a clearly identified object or of using an indicator that leads us to conclude that wars make people happy.
The main challenge that economists have to contend with, however, lies in the highly subjective nature of the phenomenon being measured. The vast majority of indicators used are constructed from satisfaction surveys, the biases of which are well-known. However, scientific studies and happiness indicators are proliferating. There is now talk of happiness economics or happiness psychology (or more broadly, positive psychology), or even more generally, the science of happiness. A scientific journal called "The Journal of Happiness Studies" was even created in 2000.
The purpose of this study is to examine the philosophical foundations of these indicators. What concept of well-being or happiness are we talking about? To what extent can well-being truly be assessed? And finally, what is the normative significance of these indicators? A thorough examination of these questions is necessary to avoid falling into the trap of measuring something without a clearly identified object or of using an indicator that leads us to conclude that wars make people happy.
Composite indexes are quantification tools based on the evaluation of the observed or projected performance or outcome of a system. They allow the aggregation of various pieces of information in order to produce synthetic and easily accessible results. There are today many composite indexes addressing political and social dimensions such as democracy or press freedom. The Prison Life Index is being elaborated and will allow the evaluation of prison conditions with regards to international standards. This tool is developed in collaboration between the LAMSADE, the LIRSA (Conservatoire national des Arts et Métiers, Paris) and Prison Insider, an information platform on prison conditions. The aim is to produce reliable and centralized information and to draw attention to the respect of the fundamental rights of prisoners in the various prisons of the world. In order to meet these objectives and provide a meaningful result methodological choices were made. The model was built collaboratively with regards to international standards on conditions of detentions. The transition from model to measurement is achieved through expert assessments seeking to overcome the limitations presented by the relatively scarce and limited quantitative data available. The aggregation model chosen to calculate the final results adopts a Multi Criteria Decision Aiding perspective (compensation levels, existence of veto thresholds). This presentation will show an applied perspective on the production of meaningful results in the context of index making.
In Multicriteria Decision Analysis, the interactions are modeled through the Shapley interaction index. To better interpret such interactions, we introuced the concept of necessary and possible interactions. We also proposed a new interaction index based on the Kemeny distance, the d-interaction index, for which the sign is always nonnegative.
18.00 - End
19.00 - Dinner at Polidor
▼
A favorite spot for Verlaine, Rimbaud, or Hemingway, and the venue for the Assemblies of the Optimates of the College of Pataphysics, the Polidor restaurant has had a unique history since its opening in 1845. Today, it is considered one of the oldest "bistros" in Paris
Program - Day 2
Session 2.1: Countering pitfalls and biases
Algorithmic decisions are now being used on a daily basis and based on Machine Learning (ML) processes that may be complex and biased. This raises several concerns given the critical impact that biased decisions may have on individuals or on society as a whole. Not only unfair outcomes affect human rights, they also undermine public trust in ML and AI. In this talk we will discuss approaches to mitigating unintended biases in ML models in order to making them fairer.
In the first part of this talk, we will address fairness issues of classifiers based on decision outcomes, and we will show how the simple idea of feature dropout followed by an ensemble approach can improve model fairness without compromising its accuracy. To illustrate we will present a general workflow that relies on explainers to tackle process fairness, which essentially measures a model's reliance on sensitive or discriminatory features. We will present different applications and empirical settings that show improvements not only with respect to process fairness but also other fairness metrics.
In the second part of this talk, we will consider the particular case of large language models and present a deep reinforcement learning approach to mitigating unintended biases such as gender, racial and demographic biases. As we will see, this task is rather complex and is far from being settled, and we will discuss its success and hidden drawbacks.
In the first part of this talk, we will address fairness issues of classifiers based on decision outcomes, and we will show how the simple idea of feature dropout followed by an ensemble approach can improve model fairness without compromising its accuracy. To illustrate we will present a general workflow that relies on explainers to tackle process fairness, which essentially measures a model's reliance on sensitive or discriminatory features. We will present different applications and empirical settings that show improvements not only with respect to process fairness but also other fairness metrics.
In the second part of this talk, we will consider the particular case of large language models and present a deep reinforcement learning approach to mitigating unintended biases such as gender, racial and demographic biases. As we will see, this task is rather complex and is far from being settled, and we will discuss its success and hidden drawbacks.
Information and information technology play a crucial role in shaping decisions across various aspects of human life. However, it is important to ensure (a) correctness, and (b) fairness in such settings. Incorrect information, such as misinformation about vaccines or elections, can lead to wrong decisions. Similarly, biased outputs from different information systems will disproportionately harm certain sections of society, such as those who identify with a specific political ideology or speak a certain language. In this talk, I will share our recent results on automatic detection of misinformation and bias, and ways to counter them. A particularly interesting scenario is one where we found misinformation detection algorithms to yield unequal performance for news articles that represented different political ideologies. I will share the key findings and potential ways to promoting correctness and fairness in such settings.
Numerous factors underlie the fact that decision support can fail to be meaningful, operational or legitimate. Power relations are among the most prominent of these factors, at two levels. On the one hand, because decision support scientists work with and for clients, and because concerned stakeholders can exert external pressures, clients and concerned stakeholders wield a power that can distort the knowledge that decision support scientists can produce. On the other hand, because decision support scientists use their knowledge to articulate advices that are expected to influence decision making, their knowledge can be seen as a form of power, and knowledge imbalances can translate into power imbalances. A vast literature explores aspects of these two kinds of power effects, but this literature is heterogeneous and often ambiguous. To clarify this picture, I propose an analysis of the concept of power, and I suggest operational criteria and procedures designed to track and prevent distortions of decision support due to power relations.
10.30 - Coffee
Session 2.2: Policies
11.00 - Robust Bidding Policies
▼
We consider a decision problem of a bidder in a market-clearing setting who has limited or no information about rivals' types or their objectives. We utilize a robust optimization approach to model limited information about rivals' behavior via an uncertainty set consisting of all possible realizations of rivals' bids. Maximizing the bidder's worst-case payoff over this set yields robust bidding policies that do not depend on distributional assumptions and are robust to information misspecification. We establish robust bidding policies in a range of auction settings, from discriminatory auctions for the unit-demand bidder to core-selecting combinatorial auctions for a single-minded and a double-minded bidder. Interestingly, this approach to modeling rivals’ information yields bidding policies that could provide a payoff that is at least as large as bidding truthfully. Furthermore, compared to expected-payoff maximizing policies, these policies could result in less allocation risk and provide higher payoff under adversarial realizations of rivals' bids.
I discuss two related objections to epistocracy that have been more or less formally formulated in the literature. The Regression Objection states that the justification of epistocracy depends on the possibility of establishing standards of expertise and competence and that this supposes that there are meta-standards identifying who is legitimate to establish them. The Selection Bias Objection contends that even if we can agree on standards of expertise and competence, biases in the selection of competent voters give no reason to believe that "competent" voters will vote more correctly than "incompetent" voters. While I argue that the Regression Objection can easily be discarded, the Selection Bias Objection is stronger. I provide reasons however to consider that it does not deliver a fatal blow to the prospects of justifying epistocratic mechanisms in liberal societies.
12.00 - Lunch
Session 2.3: Case Studies 2
I will discuss the co-construction of an operational management software for Participatory Guarantee Systems (PGS), addressing the complex challenges of organizing peer-to-peer certification visits. This software provides a "meaningful" solution by adhering to the rules explicitated by peers, making it "useful" for some PGS. However, the generalizability of its utility is debated due to specificities of PGS and need for an external actors, the developer, and its legitimacy is a subject of discussion between automation and human self-determination, central in those self-organizing peer systems.
This contribution presents a participatory decision process coordinated by the Tunisian Ministry of Agriculture to produce integrated land-use and development plans in 6 intervention zones in Tunisia. The decision process lasted over 5 years (2018-2023) and directly involved over 4,000 inhabitants in the 6 zones. Little up-to-date data existed in these zones. In this context, most of the information required for decision-making was co-constructed by the various stakeholders involved (inhabitants, researchers, administration, municipalities, etc.). Our hypothesis was that this participatory co-construction approach would increase the meaningfulness, usefulness and legitimacy of the information produced. Our contribution will retrace the participatory decision process, its successive stages, and the information that was co-produced for and by the different stakeholders at each stage. We will explain the mechanisms used to increase the meaningfulness, usefulness and legitimacy of the information produced. For example, following the identification of priority issues for the different zones, more than 11,000 proposals for action were collected from various stakeholders. These proposals responded to the local demands of the population, but were not necessarily meaningful on a regional scale. The use of "action clusters" to reflect in terms of sectors, and of matrices presenting the resources available in the area, enabled participants to question the meaningfulness, usefulness and legitimacy of the information produced. In conclusion, this experience will raise questions about the optimal way to combine stakeholder and expert information to increase the meaningfulness, usefulness and legitimacy of co-produced information in participatory decision processes.
15.00 - Coffee and discussion
16.00 - End