Theses and Dissertations from DePaul University

Date of Award

Summer 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

College of Computing and Digital Media

First Advisor

Bamshad Mobasher

Abstract

Local news outlets have experienced steep declines in readership as national and global online news sources have expanded and consolidated audience attention. Because many local media companies rely primarily on subscription-based business models, they must increase user engagement and provide clearer value to local subscribers. Personalization and recommender systems offer one promising path toward this goal. However, many recommender systems assume that a user’s past behavior can be summarized in a single unified profile. In practice, preferences are often context-dependent: the same person may prefer different content depending on geography, time, task, or situation. Local news offers a socially important example of this challenge. Readers’ local preferences often have distinctive characteristics that do not necessarily align with their broader or long-term news interests. A reader’s interest in community-relevant stories may differ from that same reader’s interest in national or international coverage, even within broad categories such as politics, sports, or culture. When all user behavior is pooled into a single representation, local interests may be diluted by broader and more popular content. Conversely, when recommendation is tuned only to local content, the reader’s broader interests may be overlooked.   To address this problem, this dissertation proposes a perspective-aware specialist-fusion machine learning framework for locality-aware recommendation. The framework models user behavior through multiple specialist views, including local, nonlocal, and topic-conditioned local perspectives, and learns to combine these views within a single deployable recommendation model. The main methodological contribution is a set of contrastive specialist-fusion approaches: Cross-Submodel Contrastive learning, which aligns different specialist representations of the same session, and Cross-Submodel Persona Contrastive learning, which incorporates persona-aware signals when user groups provide meaningful preference information. Unlike standard contrastive models that rely primarily on random sequence augmentations, the proposed methods use semantically meaningful specialist views derived from the structure of the domain and the characteristics of user preferences. This enables the model to align local, nonlocal, and topic-conditioned representations directly, rather than treating all behavior as a single pooled profile or relying on generic perturbations that may not capture meaningful preference differences.   Using real and synthetic news datasets, the dissertation shows that local and nonlocal preferences differ in measurable ways, although the strength of these differences depends on topic granularity, dataset, and domain characteristics. The experiments demonstrate that learned specialist-fusion models provide a strong foundation for integrating local and nonlocal evidence. Building on this foundation, the perspective-aware contrastive learning approach yields additional improvements in ranking accuracy, diversity, and calibration when specialist views are meaningful and sufficiently supported by the data. A study in a non-news domain further shows that the same context-conditioned modeling framework can generalize beyond local news, particularly when the relevant context is temporal rather than geographic. Overall, the dissertation demonstrates that locality-aware recommendation is most effective when local and nonlocal preferences are treated as distinct but complementary perspectives and integrated through a unified specialist-fusion model.

Download

Share

COinS