Associate team AUDITA – Data auditing systems for recommandation decision-making algorithms
Although they still remain largely unnoticed, we are today surrounded by algorithms taking decisions on our behalf. These decisions range from apparently mundane choices, such as picking a VoD movie, or selecting on-line ads, to more life-changing decisions, such the granting of a credit by a bank, the triage of patients at a hospital, or the setting of a prison term for a convicted person. In their vast majority, decision-making algorithms exploit user data to predict the likely outcome of a decision. For instance, a credit will be granted to a customer based on the likelihood that this customer will default, based on her past credit history. In spite of the pervasiveness of such decision-making algorithms, users and institutions remain largely uninformed of their precise internal workings, and in particular tend to ignore how these algorithms operate on their data . This is a fundamental societal issue, as the decisions and their explanations are most of time not provided, which lead citizens to feel confused and powerless. A decision-making algorithm essentially functions as a black-box , that consumes data collected from users (inputs), and produces decisions (outputs), while all intermediary steps remain hidden. Yet nowadays, these algorithms are executed at the service providers premises. Filter bubbles are a salient example of a problematic effect of a decision-making algorithm on users: those of recommender systems. Filter bubbles are a phenomenon where a recommendation algorithm locks the users into some narrow information bubbles with low entropy on information sources . Recommenders are then deciding which recommendations to display, while users have no understanding about the lack of diversity or the under/over-representation of particular groups of recommended items. Facing those concerns, a 2019 white paper entitled “Understanding algorithmic decision-making: Opportunities and challenges” from the European parliament, states that “Frameworks, composed of metrics, methodologies and tools that assess the impact of an Algorithmic Decision Systems and test its desired properties should be developed.” . The proposed Audita associated team aims at tackling this challenge, by the proposal of a taxonomy of feasible audit tasks, and of specific audit algorithm for recommendation systems.
Keywords: Decision making algorithm, recommendation, graph theory, machine learning models, distributed systems