Algorithmic Recourse (AR) addresses adverse outcomes in automated decision-making by ofering actionable recommendations. However, current state-of-the-art methods overlook the interdependence of features and do not consider the temporal dimension. To fill this gapt,ime-car emerges as a pioneering approach that integrates temporal information. Building upon this formulation, this work investigates the context of fairness, specifically focusing on the implications for marginalized demographic groups. Since long wait times can significantly impact communities' financial, educational, and personal lives, exploring how time-related factors afect the fair treatment of these groups is crucial to suggest potential solutions to reduce the negative efects on minority populations. Our findings set the stage for more equitable AR techniques sensitive to individual needs, ultimately fostering fairer suggestions.
Algorithmic Recourse (AR) aims to provide actionable recommendations to reverse negative outcomes from automated decision-making systems1][. Recent advancements in AR have incorporated causality to ensure plausible counterfactuals and align with human cause-efect reason2i,n3g, 4[]. However, a limitation of these methods is their inability to integrate the temporal dimension, which can lead to erroneous identification of efective features in terms of cost and time. I5n] i[s presentedtime-car, the first proposal on integrating the temporal dimension into a Causal AR problem. In this paper, we investigate the implications of fairness withinttimhe-car framework, focusing on how longer periods needed for certain tasks afect marginalized demographic groups and their connection to socioeconomic factors. The work aims to formulate fairer AR methods that are sensitive to these populations’ unique needs and time constraints.
Background Formally, the AR problem can be formulated in the following terms: given a binary classifier ℎ ∶ X → {0, 1}, and an instanc e for whichℎ( ) = 0 , the goal is to select the action∗ satisfying
∗ = arg min ( , ) .. ℎ( ( )) = 1 (1) where ( ) is a modified version of .
In [
, ) , where is the individual’s initial statei,s the target state, andis the action taken to obtain the transition between the mis. the feature space co sti,s the time cost, and is a tunable parameter that values how important is time compared to the other features. Time-unaware algorithms https://marti5ini.github.i(oM/ . Cinquini)
CEUR
ceur-ws.org
Same cost, diferent times. The temporal aspect frequently plays a crucial role in assessing the appropriateness and efectiveness of advice given to individuals by automated decision-making systems. We explore scenarios where⊂ represents sensitive attributes among the features, and we examine the case of two individuals1, and 2, with diferent sensitive attributes’ values, such t(hat) ≠ ( 2). 1 We hypothesize that the cost recommendations from the time-unaware automatic decision system for these individuals, ( ( 1), 1) and ( ( 2), 2), are approximately equal. Despite the diference in sensitive attributes, the system suggests similar cost interventions for both individuals. However, the temporal cos t could vary significantly between them, meaning one individual might need more time to achieve the desired state than the other, see1 F(Lige.ft).
Not everyone values time equally. Time may be regarded as a resource whose value varies based on individual factors. Fig1. (Right) shows this idea through applying for a loan scenario. The value of might be higher for the older population as they are likely closer to retirement and have a limited window to recover from financial setbacks. Conversely, younger individuals might have a lowevralue given their longer time horizon to adjust their savings behavior. Financial models must be calibrated to accommodate varyin g values across diferent demographic segments. This knowledge enables the design of customized recommendations sensitive to each individual’s dynamics and the time-related evaluation of changes within their specific societal and economic contexts.
Actionability as a time constraint. Actionability is considered one of the crucial aspects in a counterfactual generation proce6s]s. [We propose expanding the concept beyond the notionboeifng able to act upon to includethe ability to do so within a reasonable timeframe. Indeed, if the action required to implement a recommendation is excessively time-consuming or impractical, the recommendation becomes unhelpful for the user. In the constrained optimization framework of AR, the actionability threshold is directly controlled by the maximum time constraint, denot ed as . This parameter can be determined a priori oardapted based on the user’s specific requirements each time a request is made. In the latter case, the parameter enables personalized control over actionability for the applicant. From this perspective, we propose a new interpretatiofanirofrecommendations expressed as follows: a time-aware algorithmic recourse model is fair if its recommendations remain fair under any fixed time constraint. This work has been partially supported by the EU H2020 programme under the funding schemes ERC-2018-ADG G.A. 834756 “XAI: Science and technology for the eXplanation of AI decision making”, “SoBigData++: European Integrated Infrastructure for Social Mining and Big Data Analytics”, by the European Commission under the NextGeneration EU programme – National Recovery and Resilience Plan (Piano Nazionale di Ripresa e Resilienza, PNRR) – Project: “SoBigData.it – Strengthening the Italian RI for Social Mining and Big Data Analytics” – Prot. IR0000013 – Avviso n. 3264 del 28/12/2021, and M4C2 - Investimento 1.3, Partenariato Esteso PE00000013 - “FAIR - Future Artificial Intelligence Research” - Spoke 1 “Human-centered AI”, and by Fondo Italiano per la Scienza FIS00001966 MIMOSA.