Active deep learning on entity resolution by risk sampling

While the state-of-the-art performance on entity resolution (ER) has been achieved by deep learning, its effectiveness depends large quantities of accurately labeled training data. To alleviate data labeling burden, Active Learning (AL) presents itself as a feasible solution that focuses deemed useful for model training. Building upon recent advances in risk analysis ER, which can provide more refined estimate label misprediction than simpler classifier outputs, we propose novel AL approach sampling ER. Risk leverages estimation active instance selection. Based core-set characterization AL, theoretically derive an optimization aims to minimize loss with non-uniform Lipschitz continuity. Since defined weighted K-medoids problem is NP-hard, then present efficient heuristic algorithm. Finally, empirically verify efficacy proposed real comparative study. Our extensive experiments have shown it outperforms existing alternatives considerable margins.