Active Transfer Bagging: A New Approach for Accelerated Active Learning Acquisition of Data by Combined Transfer Learning and Bagging Based Models

TL;DR

Active Transfer Bagging (ATBagging) tackles the seed-data inefficiency in active learning by leveraging related proxy data through a two-part mechanism: (i) informativeness scores derived from a Bayesian interpretation of bagged ensembles and (ii) a determinantal point process (DPP) that enforces feature-space diversity. The method computes an information-gain proxy from in-bag vs. out-of-bag predictive distributions and builds a kernel via Random Fourier Features to sample diverse, informative seed points with a fixed size using a $k$-DPP. Across four real-world datasets representing target-transfer and feature-shift scenarios, ATBagging consistently improves early learning performance and downselection quality, with the strongest gains in low-data regimes. The work demonstrates that a principled combination of informativeness and heterogeneity can substantially reduce labeling costs while maintaining or improving predictive performance when transferring to related tasks.

Abstract

Modern machine learning has achieved remarkable success on many problems, but this success often depends on the existence of large, labeled datasets. While active learning can dramatically reduce labeling cost when annotations are expensive, early performance is frequently dominated by the initial seed set, typically chosen at random. In many applications, however, related or approximate datasets are readily available and can be leveraged to construct a better seed set. We introduce a new method for selecting the seed data set for active learning, Active-Transfer Bagging (ATBagging). ATBagging estimates the informativeness of candidate data point from a Bayesian interpretation of bagged ensemble models by comparing in-bag and out-of-bag predictive distributions from the labeled dataset, yielding an information-gain proxy. To avoid redundant selections, we impose feature-space diversity by sampling a determinantal point process (DPP) whose kernel uses Random Fourier Features and a quality-diversity factorization that incorporates the informativeness scores. This same blended method is used for selection of new data points to collect during the active learning phase. We evaluate ATBagging on four real-world datasets covering both target-transfer and feature-shift scenarios (QM9, ERA5, Forbes 2000, and Beijing PM2.5). Across seed sizes nseed = 10-100, ATBagging improves or ties early active learning and increases area under the learning-curve relative to alternative seed subset selection methodologies in almost all cases, with strongest benefits in low-data regimes. Thus, ATBagging provides a low-cost, high reward means to initiating active learning-based data collection.

Figures (19)

• Figure 1: Information flow of the knowledge gain and heterogeneity operators
• Figure 2: Overview of the three key steps in ATBagging learning approach.
• Figure 3: Parity plots demonstrating the capability of the exemplar Scikit-learn RFR model to regress the selected datasets. Each subfigure corresponds to a different dataset utilized in the numerical experiments, with target-transfer datasets appearing twice, once for each target. The left column (blue) represents the training set performance (80% of the data) and the right column (red) represents test set (20% of the data). \ref{['fig:parity:a']} the ERA5 dataset with source target total precipitation and transfer target total runoff, \ref{['fig:parity:b']} the QM9 dataset with source target LDA(VWN)-SZP energies and transfer target M06-2X-TZP energies, \ref{['fig:parity:c']} the PM2.5 dataset with target PM2.5 particle concentration, and \ref{['fig:parity:d']} the Forbes 2000 dataset with target market value.
• Figure 4: Performance of the subset selection methods for dataset downselection on the ERA5 dataset. \ref{['fig:era_down:a']} Comparison of subset accuracy for those generated by ATBagging to those of the three alternative methods. Mean performance over the replicates is shown as the solid line, with 90% high density intervals shown by the shaded regions. \ref{['fig:era_down:b']} The percentage of trials in which ATBagging outperforms the indicated alternative method. The shaded regions represent 90% credible intervals from a beta-binomial posterior of the pairwise comparison data.
• Figure 5: Performance of the subset selection methods for dataset downselection on the QM9 dataset. \ref{['fig:qm9_down:a']} Comparison of subset accuracy for those generated by ATBagging to those of the three alternative methods. Mean performance over the replicates is shown as the solid line, with 90% high density intervals shown by the shaded regions. \ref{['fig:qm9_down:b']} The percentage of trials in which ATBagging outperforms the indicated alternative method. The shaded regions represent 90% credible intervals from a beta-binomial posterior of the pairwise comparison data.