The Enforced Transfer: A Novel Transfer Learning Solution Applied to Interpersonal Conflict Detection and More
Detecting the onset of interpersonal conflict is a crucial step to help facilitate proper management. In this work, we develop a novel transfer learning solution, the Enforced Transfer, to detect interpersonal conflict in real home environments using data collected on 19 couples engaging in conversations on topics on which they disagree. The samples in this dataset are artificially augmented to account for a commonplace issue in home environments: in-home noise. The Enforced Transfer yields an F1 score of 89.93%, outperforming two state-of-the-art baselines by 14.72% and 21.08%. We also investigate if the Enforced Transfer is a generic transfer learning solution that can be applied not only to transfer-learn from the domain of emotions to the domain of conflict, but also applicable to fields such as computer vision. We compare the Enforced Transfer against eight other state-of-the-art deep transfer learning algorithms on two standard benchmarks on computer vision that deep transfer learning algorithms have been evaluated on (transfer learning from MNIST to USPS and from SVHN to MNIST). The Enforced Transfer outperforms all eight algorithms on the two tasks (with accuracy scores of 95.41% and 95.43%, outperforming the best-performing state-of-the-art algorithms by 1.31% and 1.83%). We further test the ET against five state-of-the-art baselines on a more complex transfer learning task in computer vision - to transfer-learn from CIFAR-10 to STL-10, and the Enforced Transfer achieves an accuracy of 86.12%, outperforming all the other baselines by 3.5%, 5%, 61%, 9.8%, and 19.7%. The success of the Enforced Transfer on the three computer vision transfer learning tasks suggests that our approach is a generic transfer learning solution that can be applied to not just conflict detection.
- Yangfeng Ji, Committee Chair, (CS/SEAS/UVA)
- John Stankovic, Advisor, (CS/SEAS/UVA)
- Lu Feng (CS/SEAS/UVA)
- Hongning Wang (CS/SEAS/UVA)