From Human Inputs to Actionable Insights: Expanding Data Aggregation Techniques and Models for Objective and Subjective Group Decision Making

Human-derived inputs are crucial in decision-making across various domains such as social choice and collective intelligence. These inputs, when the task is objective, can be collected as binary data, and when subjective, as preference-ordered lists. Despite their potential insights, consolidating these diverse evaluations into a unified outcome that accurately reflects group consensus is a critical challenge. This dissertation addresses the limitations of traditional aggregation techniques, introducing novel methodologies to improve aggregation quality using social choice and optimization techniques in various contexts. The first part of this dissertation develops methods to improve the quality of aggregated outcomes for objective tasks by incorporating additional participant inputs. An image labeling experiment was designed where participants classified images based on the presence of target objects. This study used inputs beyond binary classification, such as confidence values and target coordinates, within a machine learning (ML) framework to improve labeling accuracy. The results showed that these additional inputs can enhance classification accuracy, even with smaller datasets. The second part of this dissertation develops aggregation techniques for subjective tasks, addressing the misalignment with the assumption of social homogeneity. It introduces Split Consensus Ranking Aggregation (SC-RA), a social-choice inspired methodology based on distances between rankings, to identify multiple latent collective tendencies in preferences. A variant based on the Kemeny-Snell (KS) distance is also developed, featuring desirable social choice properties. Two binary programming models are proposed to solve the associated problem, which is shown to be NP-hard under certain conditions. Symmetry-breaking constraints and a data reduction technique are introduced to facilitate the solution process. The effectiveness of this methodology at recovering consensus splits is demonstrated through six real-world datasets and non-trivial synthetic data. The third part of the research involves a crowdsourcing experiment where participants provide preferences on social issues and subjective domains. The results highlight the value of capturing subgroup preferences without relying on demographic information, which is often misleading. It also demonstrates the model's effectiveness in identifying spammers who provide non-genuine responses.