Graph Transduction Techniques for Domain Adaptation. We compare our game-theoretic graph transduction technique against three other transductive techniques, namely Label Propagation (LP) [30], Label Spreading (LS) [31] and Har- monic Function (HF) [32] for the domain adaptation problem. Similar to our method, these techniques exploit the so-called smoothness principle which states that closer instances tend to belong to the same class. LP [30] performs hard clamping of input labels which yields to avoiding change on the original label distribution at every iteration, while LS [31] adopts soft clamping where initial assignments are changed by a fraction α at each iteration. Moreover, employing regularization, the cost employed in LS differentiates from LP, which provides - 5 - paper N-20296.pdf IJCNN 2019. International Joint Conference on Neural Networks. Budapest, Hungary. 14-19 July 2019 → → → → → → TABLE I: Comparative analysis on Office-31 dataset (ResNet- 50 features) A D A W D A D W W A W D avg Source SVM 76.9 73.8 60.3 97.5 59.4 100.0 78.0 Source LR 74.7 70.8 60.6 97.5 60.2 100.0 77.3 Shallow models SA 76.7 75.5 62.2 97.9 60.3 100.0 78.8 FLDA-Q 76.3 75.5 59.9 97.5 58.6 99.8 77.9 CORAL 78.9 76.9 59.7 98.2 59.9 100.0 78.9 Graph-transductive Lab Prop methods 2.4 3.6 3.3 3.6 3.3 99.8 19.3 Lab Spread 77.3 79.2 63.1 98.6 60.8 99.8 79.8 Baselines Harmonic Function 73.7 80.3 62.3 98.1 46.8 99.8 76.8 Proposed method, GTDA GTDA 80.5 78.0 66.2 98.9 62.9 99.8 81.1 GTDA + LR 82.5 84.2 67.1 97.9 69.1 99.8 83.4 → → → → → → TABLE II: Comparative analysis on Office-31 dataset (ResNet-50 features) A D A W D A D W W A W D avg Deep Neural Networks (results taken from [16]) DDC 76.5 75.6 62.2 96.0 61.5 98.2 78.3 XXX 78.6 80.5 63.6 97.1 62.8 99.6 80.4 RTN 77.5 84.5 63.6 96.8 64.8 99.4 81.6 RevGrad 79.7 82.0 68.2 96.9 67.4 99.1 82.2 JAN-A 85.1 86.0 69.2 96.7 70.7 99.7 84.6 Proposed method, GTDA GTDA 80.5 78.0 66.2 98.9 62.9 99.8 81.1 GTDA + LR 82.5 84.2 67.1 97.9 69.1 99.8 83.4 better robustness to noise. Gaussian Fields and Harmonic Functions (HF) [32] tries instead to compute a function f by minimizing a corresponding energy function E(f ). The solution is harmonic and this property can be exploited to propagate information according to the aforementioned smoothness principle.
