Research Themes

Retrieval-Enhanced Machine Learning (REML) explores how machine learning models can be augmented with retrieval components to improve performance, interpretability, and scalability. This framework treats information access systems as supporting both human users and task-driven machines, applying cor...

This research area examines how AI systems, particularly recommender systems and language models, shape and interact with cultural content and experiences. This includes developing metrics like commonality to measure how recommendation systems promote shared cultural experiences across populations, ...

This research area focuses on core methods for evaluating AI systems, including development of quantitative metrics, mixed methods approaches that combine qualitative with quantitative insights, and contextual meta-evaluation methods.

Tip-of-the-tongue (TOT) retrieval addresses the challenge of helping users find specific information like movies or songs when they cannot recall specific identifiers like titles. People express information needs during TOT states using combinations of content descriptions, contextual memories, and ...

This research area addresses the broader implications of AI systems, with particular focus on multi-stakeholder environments and inter-group fairness. Key work includes developing frameworks for auditing search engines across groups, creating metrics that balance consumer and producer interests in ...

This research develops evaluation methods that move beyond traditional scalar metrics to directly model user preferences and system robustness. Key contributions include recall-paired preference (RPP), a metric-free evaluation method that computes preferences between ranked lists while simulating mu...

This research addresses the complex dynamics of systems serving multiple stakeholders with potentially conflicting interests, such as consumers and producers in marketplaces. Key contributions include developing relevance and ranking algorithms for online dating systems that consider two-sided relev...

Behavior modeling focuses on understanding and predicting user interactions with AI systems through behavioral signals and interaction patterns. Key work includes developing models for search result prefetching using cursor movements and viewport behavior, studying mobile query reformulation pattern...

This research area explores how to represent and protect user information in personalized systems while maintaining privacy and transparency. Key work includes developing frameworks for data minimization that learn to limit personal data collection based on performance curves and legal principles li...

Crisis informatics research develops computational methods for processing and analyzing information during mass emergencies and disasters. This work addresses challenges including parsing informal messages, handling information overload, and prioritizing actionable information for first responders a...

Temporal dynamics research examines how time affects information retrieval and ranking, focusing on the temporal characteristics of queries, documents, and user information needs. This work includes developing temporal profiles of queries to predict retrieval performance, creating recency-sensitive ...

Aggregated search focuses on integrating results from multiple specialized search engines (verticals) into unified result presentations, addressing both vertical selection and result placement decisions. This research area encompasses predicting which verticals are relevant for queries, learning mod...

This research area applies machine learning techniques to improve various aspects of information retrieval, from ranking and matching to cross-market adaptation and personalization. Key work includes developing neural ranking models that combine local and distributed text representations, creating d...

Pseudo-relevance feedback research develops methods for automatic query expansion and retrieval improvement without explicit user feedback. Key contributions include condensed list relevance models that provide nearly identical performance to re-retrieval with significantly lower latency, and pseudo...

Score regularization is based on the principle that topically related documents should receive similar retrieval scores, extending the cluster hypothesis to score-based information retrieval. This research demonstrates that adjusting retrieval scores to ensure local consistency (autocorrelation) sig...

Query performance prediction research develops methods to estimate retrieval effectiveness without requiring user feedback or relevance judgments. Performance predictors can be effectively applied to specialized domains like speech-enabled search, helping to select optimal query transcriptions and d...