Adaptive deep learning for entity resolution by risk analysis

The state-of-the-art performance on entity resolution (ER) has been achieved by deep learning. However, models usually need to be trained large quantities of accurately labeled training data, and cannot easily tuned towards a target workload. in real scenarios, there may not sufficient data; even if they are abundant, their distribution is almost certainly different from data some extent. To alleviate such limitation, this paper proposes novel risk-based adaptive approach for ER that can tune model its workload the workload’s particular characteristics. Built recent advances risk analysis ER, proposed first trains then fine-tunes it unlabeled minimizing misprediction risk. Our theoretical shows correct label status mispredicted instance with fairly good chance. Finally, we empirically validate efficacy benchmark comparative study. extensive experiments show considerably improve models. Furthermore, scenario misalignment, similarly outperform alternatives transfer learning considerable margins. Using as test case, demonstrate promising potentially applicable various challenging classification tasks.