Document Set Expansion with Positive-Unlabelled Learning Using Intractable Density Estimation

TL;DR

The paper addresses Document Set Expansion (DSE) under Positive-Unlabeled (PU) learning without access to the positive class prior. It introduces puDE, a framework that leverages intractable density estimation to learn a Bayesian classifier by estimating densities f_P and f with two models, yielding ${P}(Y=1|\mathbf{x}) \approx \frac{p_\theta(\mathbf{x})}{q_\theta(\mathbf{x})}\,\pi$; it implements two branches: KDE (nonparametric) and Energy-Based Models (parametric), both trained via maximum likelihood with MCMC/Langevin dynamics in a transductive setting. The authors validate puDE on PubMed-derived topics and a Covid-19 literature dataset, showing superior F1 and ranking metrics compared with BM25, nnPU, and VPU, and demonstrate robustness to small labelled sets. The work provides a practical PU-learning solution for real-world DSE, with code available at the cited GitHub repository.

Abstract

The Document Set Expansion (DSE) task involves identifying relevant documents from large collections based on a limited set of example documents. Previous research has highlighted Positive and Unlabeled (PU) learning as a promising approach for this task. However, most PU methods rely on the unrealistic assumption of knowing the class prior for positive samples in the collection. To address this limitation, this paper introduces a novel PU learning framework that utilizes intractable density estimation models. Experiments conducted on PubMed and Covid datasets in a transductive setting showcase the effectiveness of the proposed method for DSE. Code is available from https://github.com/Beautifuldog01/Document-set-expansion-puDE.

Figures (2)

• Figure 1: Inductive setting and transductive setting in PU learning.
• Figure 2: F1 comparison on covid dataset with respect to the ratio of |LP| over |U| ranging from 0.01 to 0.1 with step of 0.01 and from 0.1 to 1 with step of 0.1.