Date of Award

2026

Document Type

Thesis (Ph.D.)

Department or Program

Computer Science

First Advisor

Soroush Vosoughi

Abstract

Intelligent systems increasingly anticipate which items a person will prefer, adapt generated text to a familiar style, and carry relevant information across interactions. These capabilities invite the impression that a model knows its user. In the technical sense studied here, a model knows a user when it carries information specific to that user, uses that information in the relevant context, and generalizes beyond the observations used to build the user state. Knowing is thus a measurable relation among user evidence, model state, context, and held-out user-specific behavior. The research community generally studies this family of systems under the term personalization, which broadly encompasses methods that use user information to adapt predictions or outputs. Personalization and knowing, however, make different claims. A profile, interaction history, or improvement on a personalized task may establish that user information was available or useful, but it does not establish what the model retained, whether that information supported the current decision, or whether its influence was specific to the target user.

To address this evidentiary gap, this dissertation presents six studies spanning recommendation and personalized language modeling that examine the path from observation to prediction through representation, contextualization, and predictive internalization. Representation asks what user information learning preserves. The representation studies show that preserving meaningful preference distinctions determines whether the learned user state can support generalization. Preservation alone is insufficient because the same information should not influence every prediction uniformly. Contextualization therefore asks when preserved user information should affect a decision. The contextualization studies show that the value of past behavior depends on the current item or task, the structure of observed interactions, and the strength and temporal relevance of the expressed preference. Contextualization thus requires selective use of user evidence rather than the accumulation of more history. Even relevant user evidence may not influence prediction. Predictive internalization therefore asks whether user evidence changes predictions of held-out behavior in a user-specific way. The predictive-internalization study shows that this influence becomes observable when target-user evidence improves prediction of held-out behavior and the improvement follows the correct user rather than surface similarities or generic context.

Together, the three perspectives establish knowing the user as an evidentiary connection from observed behavior to user-specific prediction. This account directs evaluation toward the evidence a model preserves, the conditions governing its use, and the predictive changes attributable to the target user. The same account also exposes three boundaries for personalized models. Personal information requires control over where its influence persists and which contexts authorize its use. The relevance of past behavior varies across tasks and over time. Behavioral predictability leaves the user's underlying intentions uncertain. These boundaries motivate models whose user knowledge is generalizable, contextually relevant, empirically observable, and governed by the person it describes. By connecting these requirements and boundaries, this dissertation situates personalization within the broader question of how intelligent systems form, apply, and govern knowledge about people. It advances a view of knowing the user as an empirically grounded and accountable relationship between models and the individuals they represent.

Download

Share

COinS