Periodic Reporting for period 3 - HumanML (Human-Centric Machine Learning)
Reporting period: 2023-11-01 to 2025-04-30
With the advent of mass-scale digitization of information and virtually limitless computational power, an increasing number of social, information and cyber-physical systems evaluate, support or even replace human decisions using machine learning models and algorithms. However, as these decisions become more consequential to individuals and society, machine learning models and algorithms have been blamed to play a major role in an increasing number of missteps, from discriminating minorities, causing car accidents and increasing polarization to misleading people in social media. What went wrong in these cases? In most social, information and cyber-physical systems, algorithmic and human decisions feed on and influence each other, however, machine learning models and algorithms have been traditionally designed to take decisions autonomously, without human intervention, on the basis of passively collected data. In a way, to replace humans.
In this project, the overall objective is to develop human-centric machine learning models and algorithms for evaluating, supporting and enhancing decision making processes where algorithmic and human decisions feed on and influence each other. These models and algorithms account for the feedback loop between algorithmic and human decisions, which currently perpetuates or even amplifies biases and inequalities, and they learn to operate under different automation levels. Moreover, they anticipate how individuals will react to their algorithmic decisions, often strategically, to receive beneficial decisions and provide actionable insights about their algorithmic decisions. Finally, these models and algorithms are useful in a wide range of applications in health, justice and recruiting, to name a few.
In this project, the overall objective is to develop human-centric machine learning models and algorithms for evaluating, supporting and enhancing decision making processes where algorithmic and human decisions feed on and influence each other. These models and algorithms account for the feedback loop between algorithmic and human decisions, which currently perpetuates or even amplifies biases and inequalities, and they learn to operate under different automation levels. Moreover, they anticipate how individuals will react to their algorithmic decisions, often strategically, to receive beneficial decisions and provide actionable insights about their algorithmic decisions. Finally, these models and algorithms are useful in a wide range of applications in health, justice and recruiting, to name a few.
We have developed machine learning models that are optimized to operate under different automation levels, i.e. these models are optimized to take a given fraction of the decisions and leave the remaining ones to humans. We have developed models for two distinctive settings: (i) decisions are independent and there are multiple human decision makers (e.g. medical diagnosis); and, (ii) decisions are dependent and take place over time and there is a single human decision maker (e.g. semi-autonomous driving).
We have developed machine learning models for decision making whenever decision subjects behave strategically in order to receive beneficial decisions. Our models incentivize individuals to invest in forms of effort that increase the utility and fairness of the decisions and the allow both for settings in which the decision maker wishes to be fully transparent about the machine learning model it uses to take decisions and settings in which the decision maker wishes to be partially transparent and only reveal part of the machine learning model it uses.
We have developed machine learning methods to facilitate counterfactual thinking, the process of mentally simulating alternative worlds where events of the past play out differently than they did in reality. Our method focus on decision making processes in which multiple, dependent actions are taken sequentially over time. More specifically, they find alternative sequence of actions differing in a few actions from the observed sequence that could have led the observed process to a better outcome.
We have developed a machine learning algorithm to help people memorize more effectively and, together with the startup Swift Management AG, carried out a large-scale randomized control trial to show its effectiveness in a real e-learning app.
In the context of COVID-19, we have developed a machine learning algorithm to perform more efficient pooled testing.
We have developed machine learning models for decision making whenever decision subjects behave strategically in order to receive beneficial decisions. Our models incentivize individuals to invest in forms of effort that increase the utility and fairness of the decisions and the allow both for settings in which the decision maker wishes to be fully transparent about the machine learning model it uses to take decisions and settings in which the decision maker wishes to be partially transparent and only reveal part of the machine learning model it uses.
We have developed machine learning methods to facilitate counterfactual thinking, the process of mentally simulating alternative worlds where events of the past play out differently than they did in reality. Our method focus on decision making processes in which multiple, dependent actions are taken sequentially over time. More specifically, they find alternative sequence of actions differing in a few actions from the observed sequence that could have led the observed process to a better outcome.
We have developed a machine learning algorithm to help people memorize more effectively and, together with the startup Swift Management AG, carried out a large-scale randomized control trial to show its effectiveness in a real e-learning app.
In the context of COVID-19, we have developed a machine learning algorithm to perform more efficient pooled testing.
The project has advanced the state of the art in (i) machine learning models that operate under different automation levels, (ii) machine learning models that operate under strategic behavior, and (iii) machine learning models that facilitate counterfactual thinking. Until the end of the project, we plan to further advance the state of the art in (i-iii).
In regards to (i), we will develop machine learning models that simplify decision making tasks in high-stakes domains where the expert takes the final decision for every decision instance. These models leverages intuitive, counterfactual properties about human behavior to help experts take better decisions.
In regards to (ii), we will develop machine learning models to facilitate contesting by humans subjects in algorithmic decision making.
In regards to (iii), we will develop machine learning models to better understand how human experts use counterfactual thinking for decision making.
In regards to (i), we will develop machine learning models that simplify decision making tasks in high-stakes domains where the expert takes the final decision for every decision instance. These models leverages intuitive, counterfactual properties about human behavior to help experts take better decisions.
In regards to (ii), we will develop machine learning models to facilitate contesting by humans subjects in algorithmic decision making.
In regards to (iii), we will develop machine learning models to better understand how human experts use counterfactual thinking for decision making.