Schedule

Algorithmic fairness is a key requirement of algorithmic decision-making systems. Over the last decade, a large body of research has shown that models developed without attention to protected categories harm vulnerable populations when deployed in societal contexts. Fieldwork and legal analysis highlight an often overlooked tension between data protection principles and algorithmic fairness methods that require access to individual sensitive attributes. In this talk, I argue that fairness cannot be controlled unless reliably estimated from realistic measurements. I will discuss different approaches to measuring fairness under credible data limitations on sensitive attributes safeguarding individual privacy and aligning with data protection.

The opinion dynamic within networked systems is significantly shaped by individual and social factors, acting as either catalysts or obstacles to the acceptance of innovative ideas or products. This phenomenon is evident in the examination of open-loop dynamics characterizing the social environment and becomes equally vital when devising promotion policies aimed at accelerating the diffusion of innovations within it.  Indeed, the success of such policies is intricately tied to the distinctive socio-economic and epistemic (i.e., knowledge-related) characteristics of the environment in which they are implemented. Therefore, these aspects must be carefully considered in the decision-making process to guarantee effective fostering plans. 

By focusing on the diffusion of sustainable technologies crucial for meeting environmental goals in response to the current climate emergency, our proposal introduces a comprehensive framework for addressing societal challenges within the realm of innovation diffusion. Encompassing all essential steps such as analysis, modeling, policy design and simulation, our framework aims to facilitate a fair and effective diffusion of sustainable technologies. Specifically, we propose a data-driven approach that integrates socio-economic and epistemic factors into opinion dynamic models to study the effect of individual attitudes and social imitation (or discrimination) on the dynamics of opinion diffusion. These data-based models are leveraged to design effective promotion policies aiming at maximizing the widespread adoption of innovative (and sustainable) ideas or products, maintaining policy costs at a minimum. Moreover, in directly addressing societal aspects, our policy design framework aims to actively incorporate social justice and inclusion objectives into the optimal control scheme to ensure a fair innovation diffusion. We finally exploit a simulation-based methodology to assess and compare the impact of such multi-objective policies. This holistic approach deepens our understanding of the importance of fairness goals in the dissemination of innovative technologies and their intricate relationship with the traditional concept of performance in control systems.

Global urbanization and burgeoning urban populations impose several societal challenges associated with disparities in transportation opportunities, reduced accessibility to essential services for marginalized communities, and increased social isolation due to lengthy commutes. Although emerging mobility systems, e.g., connected and automated vehicles (CAVs), and shared mobility, provide the most intriguing opportunity to mitigate these challenges, research efforts have focused mainly on optimizing their operation efficiency in isolation without deliberating on human acceptance and perception. Addressing the societal issues inherent in emerging mobility systems remains largely uncharted territory. The core of these societal challenges lies in the unequal distribution of transportation modes and access to urban resources, giving rise to "mobility equity." While mobility equity has been studied from multiple perspectives, including socioeconomic parity, equitable spatial infrastructure allocation, and alignment of resource distribution with societal needs, a critical gap remains in integrating mobility equity principles into emerging transportation modes. In this talk, I will present a mobility equity metric (MEM) to quantify the accessibility and fairness in a transportation network consisting of CAVs and human-driven vehicles. I will then present a decision-making framework integrated with MEM to distribute travel demand for the transportation network, resulting in a socially optimal mobility system. A “socially optimal mobility system” is defined as a mobility system that (1) is efficient in terms of travel time, (2) improves accessibility, and (3) ensures equity in transportation. The framework incorporates a cognitive hierarchy model commonly used in behavioral economics to predict human decisions in transportation systems to accommodate compliant and noncompliant vehicles to the routing suggestions. The proposed framework aims to bolster mobility equity by addressing transportation and resource access disparities.

The simultaneous expansion of electric vehicles (EVs) and intermittent renewable energy sources holds the potential to accelerate the decarbonization of highly emissions-intensive sectors. However, these developments also pose challenges to the stability of power systems, which could impede their widespread adoption. Nevertheless, electric vehicles present a significant opportunity to enhance flexibility, enabling better integration of renewable energy variations and optimization of energy markets. To address these challenges, there is a need to design new models that incorporate both electric vehicle traffic flows and battery charge dynamics. These models represent a crucial step towards utilizing e-flexibility, a concept aimed at predicting the spatial and temporal evolution of EVs' state of charge in relation to the electrical grid and associated electricity market operations. The presentation introduces a graph-based approach to model the mobility of electric vehicles and the evolution of their state of charge in large-scale urban traffic networks. This model combines the vehicles' mobility, described by dynamic equations over a graph that captures origin-destination movements, with the energy consumption associated with their mobility patterns. The model also incorporates power inputs from charging stations. We will showcase the operation of these models using our numerical twin (eMob-Twin) on the open-to-public platform emob-twin.inrialpes.fr. This model can be extended to account for driver behavior in determining when and where to charge, considering factors such as current state of charge, distance to charging stations, and charging costs. Moreover, it can be used to identify optimal locations for charging stations, maximizing convenience for EV users and profitability for charging station owners. We will also introduce a simplified (averaged) version of the model, which serves as a foundation for optimizing charging station locations while reducing computational complexity. Lastly, we propose an innovative approach that utilizes aggregated EVs for grid-balancing services in the auxiliary market.  This is done by an optimization framework which establishes pricing strategies to maximize profits for aggregators and CSOs while minimizing charging costs for EV users. Our findings demonstrate the effectiveness of this strategy in realistic simulations, integrating EV mobility and the Electricity FCR market.

Risk-aversion plays a crucial role in dealing with uncertainty in many societal-scale networks such as online marketing and transportation systems. This talk will explore managing uncertainty through risk-averse decision-making in online convex games, where multiple agents aim to minimize their Conditional Value at Risk (CVaR) with different risk levels. Recognizing that agents typically possess varied information in practical situations, we examine three distinct game setups and design algorithms for each one of them. In the first risk-averse game, all agents rely solely on zeroth-order oracles, that is, each agent receives the cost function value after each iteration. In this case, we show that each agent achieves no-regret learning with high probability. In the second game, the agents instead have access to first-order gradient evaluations of the cost function, which again leads to no-regret learning. Under additional assumptions on the cost function, the converges to a Nash equilibrium with high probability is guaranteed. The third game is a hybrid setting with agents accessing either zeroth-order oracles or stochastic gradients. For this setup, we design a Nash equilibrium-seeking algorithm for each agent depending on its feedback information. It is shown that our algorithm always performs between pure zeroth-order and first-order methods. Throughout the presentation we illustrate the results through the classical Cournot game in economics.

Social media platforms employ recommender systems to engage users with content. Recommender systems, while powerful tools for enhancing user experience and content consumption, have undoubtedly contributed to the emergence of echo chambers over digital platforms and consequently to undesired effects over users’ opinion, such as polarization. In order to limit these undesired consequences and it is of paramount importance to give to this phenomenon a quantitative slant.

This talk develops along a quite unexplored perspective on recommender systems where feedback interaction between users and recommender systems is made explicit. User interactions, real-time opinions and user clicking behavior are assumed unknown for online recommendations.

Our recommender scheme shows promising performances and allows to independently test the effect of users’ clicking behavior, social network topology and objective function of the recommender systems over the opinion dynamics.

I will discuss a model of opinion dynamics over networks that reveals the fundamental roles of nonlinearity, negative and positive feedback, and network structure in opinion formation, opinion cascades, and decision-making. The model represents a vast range of possible dynamic network behaviors, while retaining analytical tractability. Thus, it provides a principled and systematic means for exploring open questions, deriving and evaluating hypotheses, and advancing design for fair decision-making and societally-aware control in networked systems. I will present key theoretical results and illustrate their utility in the context of the workshop, for example, to study how the structure of a social interaction network affects cooperative behavior in social dilemmas like the public goods game.

We study opinion dynamics in social settings involving agents on a network who are affected by social pressure and may not disclose what they really think. Under these dynamics, we analyze how to determine the long-term behavior of agents declared opinions and more importantly,  how to effectively infer their true beliefs. The talk stems from a joint work with Jennifer Tang and Amir Ajorlou.

Dynamic models of social processes on networks are powerful tools for studying emergent collective behaviour, including for designing control policies and interventions that are fair and societally-aware. However, many models rely on assumptions about social influence and decision-making mechanisms that are not grounded by empirical evidence, and identifying model parameters associated with real-world situations can be challenging. This means that applying such models to real-world situations can be challenging, and picking inappropriate assumptions or parameter values will lead to incorrect predictions from the model. In this work, I will review two recent efforts to resolve these issues, showing how carefully designed experiments with human experiments can be used examine model assumptions and estimate model parameters, and highlight the future need to closely integrate experimental and modelling efforts for studying social dynamics.                               

In the first problem, we consider how social conventions form and evolve. Our experiments show that, besides the standard assumption of coordination, individuals are also influenced by inertia (a tendency to stick to their current decision) and dynamic norms (a tendency to be influenced by trends). Incorporating these mechanisms into an existing model greatly improves the model’s fit to the experimental data, and allows the model to generate a richer set of collective dynamics. In the second problem, we consider how information spreads through online social media. Our experiments show that social influence plays a minimal role in determining information transmission, in contrast to assumptions that underpin epidemic and complex contagion models for information spread. Rather, message content appears to be the key determining factor for an individual to transmit information, while viral spread may be linked to platform affordances, such as the recommender system. Our findings have implications in dealing with misinformation spread on social media.