Unsupervised Bootstrapping of Active Learning for Entity Resolution

Abstract

Entity resolution is one of the central challenges when integrating data from large numbers of data sources. Active learning for entity resolution aims to learn high-quality matching models while minimizing the human labeling effort by selecting only the most informative record pairs for labeling. Most active learning methods proposed so far, start with an empty set of labeled record pairs and iteratively improve the prediction quality of a classification model by asking for new labels. The absence of adequate labeled data in the early active learning iterations leads to unstable models of low quality which is known as the cold start problem. In our work we solve the cold start problem using an unsupervised matching method to bootstrap active learning. We implement a thresholding heuristic that considers pre-calculated similarity scores and assigns matching labels with some degree of noise at no manual labeling cost. The noisy labels are used for initializing the active learning process and throughout the whole active learning cycle for model learning and query selection. We evaluate our pipeline with six datasets from three different entity resolution settings using active learning with a committee-based query strategy and show it successfully deals with the cold start problem. Comparing our method against two active learning baselines without bootstrapping, we show that it can additionally lead to overall improved learned models in terms of F1 score and stability.

Deep Indexed Active Learning for Matching Heterogeneous Entity Representations

“…The latter uses red-blue set cover to learn an effective blocking function but its need for labeled data makes it ineffective in settings that call for active learning. While other approaches for blocking are available (see [11] for an exhaustive list), a number of these utilize unsupervised learning [1,39]. While unsupervised learning of a blocking function may be applicable to our setting, it does not take advantage of the labels provided by the user and thus may not adapt as well to the data at hand compared to an approach such as DIAL which does take advantage of newly labeled pairs.…”

Active Learning Based Similarity Filtering for Efficient and Effective Record Linkage

“…Even though crowdsourcing has been explored for record linkage [23] to mitigate the lack of ground truth data, allowing the public to classify record pairs is not applicable in many domains due to privacy concerns [4]. Active learning approaches, where a small number of selected record pairs are manually classified by trusted domain experts, have therefore been adopted for record linkage to generate ground truth data suitable to train supervised classifiers [17,18,24], or to generate high quality blocking results [21]. Active learning based on domain expertise, while being able to generate high quality ground truth data, can however only generate small numbers of labelled record pairs.…”

Addressing practical challenges in Active Learning via a hybrid query strategy

“…Gong et al [14] propose a novel inference method based on Bayesian Deep Latent Gaussian Model (BELGAM) to select initial training instances. Another option to address the cold-start problem is an unsupervised matching method that proposes bootstrapping active learning [21]. Furthermore, Deng et al [9] introduced a sequence-based adversarial learning model to select initial set of training instances for the AL methods.…”