Jeff Pasternack

4Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Jeff Pasternack focuses on information extraction, computational trust, and fact-finding algorithms for web data. His early work addressed the structural extraction of article text from noisy web pages before shifting toward the evaluation of source credibility and conflicting claims. His most recent research investigates latent credibility analysis to improve the transparency and accuracy of automated trust decisions.

Research Trajectory

Pasternack's early research addressed the foundational challenge of isolating meaningful content from noisy web environments. In Extracting article text from the web with maximum subsequence segmentation, he developed an approach to separate core article text from uninformative elements like navigation menus, user comments, and advertisements. As the primary author, he proposed a method that bypassed the limitations of existing techniques—such as brittle manual rules, wrapper induction requiring labeled examples for every page structure, and template detection—to more robustly segment web documents.

Following this work on information extraction, Pasternack's focus shifted to the subsequent problem of information credibility: determining which extracted claims to believe when multiple web sources disagree. In Generalized fact-finding, he introduced a framework that expanded upon traditional fact-finding algorithms. While previous state-of-the-art methods operated similarly to Kleinberg's Hubs and Authorities—iteratively computing source trustworthiness and claim believability based solely on bipartite "who claims what" relationships—his generalized approach incorporated critical contextual details. He further expanded on this framework in Making Better Informed Trust Decisions with Generalized Fact-Finding, demonstrating how integrating background knowledge, extraction uncertainty, source attributes, and claim similarity improves automated trust decisions. This body of work contributes to the broader research area of Reputation Bootstrapping for New Network Participants, addressing how decentralized systems and algorithms establish trust and evaluate claims from various unverified sources.

In 2013, Pasternack continued to refine computational approaches to web credibility. In Latent credibility analysis, he addressed the limitations of existing transitive voting systems used in fact-finding. When dealing with diverse data gathered from sources ranging from booksellers to Wikipedia pages and stock analyst predictions, traditional fact-finders often lack transparency, obfuscating how they arrive at their credibility decisions. Pasternack's work tackled this opacity, proposing latent credibility analysis to make automated credibility decisions more interpretable and robust when resolving mutually exclusive claims.

Current Directions

The most recent paper in the provided evidence is from 2013. Based on this work, Pasternack's latest documented research directions center on improving the transparency and interpretability of information credibility algorithms. Rather than relying on opaque transitive voting systems, this work investigates latent variable models to resolve conflicting claims from multiple web sources. The primary focus is on ensuring that automated trust decisions—whether applied to encyclopedic knowledge, financial predictions, or e-commerce data—are both accurate and structurally explainable.

Profile updated

See something wrong? Report a mistake