Optimal multi-attribute decision making in social choice problems

Abstract

Direction 2: Learning Preferences From Data. I model preferences using representation schema inspired by work on lexicographic heuristic decision involving multiple factors from the psychology literature, as well as using machine learning techniques. By developing novel natural language features to describe content, and using models of decision making from psychology literature that require very little training data, I find that: We can model the factors affecting voting behavior on real world data from social media and laboratory sourced question answering data with applications to crowdsourcing.

This thesis addresses problems of collective decision making when the alternatives are characterized by multiple attributes, when the agents have preferences over the alternatives. Such social choice problems arise commonly in the allocation of different types of resources in cloud computing, the allocation and exchange of medical resources, and voting over multiple ballot measures. Social media platforms often rely on voting by users to sort the deluge of content. Due to the emergence of social media as an important medium for sharing information and conducting debates, it has become increasingly important to model and understand users' voting behavior which may be affected by several attributes of the content. The problem of finding the best decision becomes challenging because (i) the number of alternatives grows exponentially with the number of attributes, and (ii) agents' preferences over the alternatives may have a complex combinatorial structure.

When alternatives are characterized by multiple attributes, several positive results in decision making literature disappear and computing optimal decisions is often hard in general. Due to the cognitive demands of forming and communicating preferences over an exponentially large number of alternatives, eliciting agents' preferences over all alternatives directly is often impractical. Driven by these concerns, several natural and simple structures for representing preferences over the values of multiple attributes have been developed in the cognitive science and artificial intelligence literature. These structures on agents' preferences make several important problems in social choice tractable by acting as a restriction on the problem domain. In this thesis, we identify cases where such structure accurately models preferences, and provide efficient mechanisms to compute optimal decisions for important social choice decision making problems with theoretical guarantees.

I study decision making problems when the alternatives are characterized by multiple attributes in two directions:

Direction 1: Decision Making Under Preferences. I consider the problem of decision making under preferences, where agents have preferences over the alternatives, and the goal is to find the best decision for the agents. For example, in cloud computing, the goal is to find the best way to allocate resources of different types to agents. When there are multiple ballot measures, the goal is to find the best decision on each ballot measures. The main theme is the identification of reasonable restrictions on the problem domain under which positive results may be recovered. Specifically, I build on a large and growing literature in social science and artificial intelligence, which suggests that agents' preferences often have some natural structure. This makes several important problems tractable. This thesis identifies cases where such structure accurately models preferences in real world data, and provides efficient mechanisms to compute optimal outcomes for important social choice problems with theoretical guarantees. This thesis delivers the following positive message: We can design efficient mechanisms with desirable properties for multi-attribute decision making under natural assumptions on agents’ preferences.