ATTENTIVE RECURRENT ADVERSARIAL DOMAIN ADAPTATION WITH TOP-K PSEUDO-LABELING FOR TIME SERIES CLASSIFICATION

摘要

THE KEY CHALLENGE OF UNSUPERVISED DOMAIN ADAPTATION (UDA) FOR ANALYZING TIME SERIES DATA IS TO LEARN DOMAIN-INVARIANT REPRESENTATIONS BY CAPTURING COMPLEX TEMPORAL DEPENDENCIES. IN ADDITION, EXISTING UNSUPERVISED DOMAIN ADAPTATION METHODS FOR TIME SERIES DATA ARE DESIGNED TO ALIGN MARGINAL DISTRIBUTION BETWEEN SOURCE AND TARGET DOMAINS. HOWEVER, EXISTING UDA METHODS (E.G. R-DANN PURUSHOTHAM ET AL. (2017), VRADA PURUSHOTHAM ET AL. (2017), CODATS WILSON ET AL. (2020)) NEGLECT THE CONDITIONAL DISTRIBUTION DISCREPANCY BETWEEN TWO DOMAINS, LEADING TO MISCLASSIFICATION OF THE TARGET DOMAIN. THEREFORE, TO LEARN DOMAIN-INVARIANT REPRESENTATIONS BY CAPTURING THE TEMPORAL DEPENDENCIES AND TO REDUCE THE CONDITIONAL DISTRIBUTION DISCREPANCY BETWEEN TWO DOMAINS, A NOVEL ATTENTIVE RECURRENT ADVERSARIAL DOMAIN ADAPTATION WITH TOP-K TIME SERIES PSEUDO-LABELING METHOD CALLED ARADA-TK IS PROPOSED IN THIS PAPER. IN THE EXPERIMENTS, OUR PROPOSED METHOD WAS COMPARED WITH THE STATE-OF-THE-ART UDA METHODS (R-DANN, VRADA AND CODATS). EXPERIMENTAL RESULTS ON FOUR BENCHMARK DATASETS REVEALED THAT ARADA-TK ACHIEVES SUPERIOR CLASSIFICATION ACCURACY WHEN IT IS COMPARED TO THE COMPETING METHODS.