Unsupervised Domain Adaptation using Graph Transduction Games

Unsupervised domain adaptation (UDA) amounts to assigning class labels to the unlabeled instances of a dataset from a target domain, using labeled instances of a dataset from a related source domain. In this paper, we propose to cast this problem in a game-theoretic setting as a non-cooperative game and introduce a fully automatized iterative algorithm for UDA based on graph transduction games (GTG). The main advantages of this approach are its principled foundation, guaranteed termination of the iterative algorithms to a Nash equilibrium (which corresponds to a consistent labeling condition) and soft labels quantifying the uncertainty of the label assignment process. We also investigate the beneficial effect of using pseudo-labels from linear classifiers to initialize the iterative process. The performance of the resulting methods is assessed on publicly available object recognition benchmark datasets involving both shallow and deep features. Results of experiments demonstrate the suitability of the proposed game-theoretic approach for solving UDA tasks.

💡 Research Summary

The paper addresses unsupervised domain adaptation (UDA) by casting the label‑propagation problem as a non‑cooperative multiplayer game on a graph, known as Graph Transduction Games (GTG). Each data point—whether from the labeled source domain or the unlabeled target domain—is treated as a player that selects a strategy corresponding to one of the class labels. Players interact through a weighted undirected graph where edge weights encode similarity between feature vectors. The core of the method is a mixed‑strategy profile: each player holds a probability distribution over the label set, reflecting uncertainty. The dynamics of the game are driven by discrete Replicator Dynamics, which iteratively update each player’s mixed strategy based on the payoff received from its neighbors. The payoff matrix for a pair of players is a scaled identity matrix whose scaling factor is the similarity weight, ensuring that the more similar two instances are, the stronger the influence they exert on each other’s label distribution.

The algorithm proceeds as follows: (1) source and target features are jointly normalized (either by standard deviation scaling or z‑score). (2) An affinity matrix W is built using a Gaussian kernel on cosine distances, with a locally adaptive bandwidth σ_i set to the distance of the 7‑th nearest neighbor of each point. (3) To guarantee graph connectivity, a sparsification step retains only the k = ⌈log₂(N)⌉+1 nearest neighbors for each node, symmetrizing the adjacency. (4) The mixed‑strategy profile is initialized either uniformly (no prior) or with soft pseudo‑labels obtained from a logistic regression classifier trained on the source data (the “+LR” variant). Logistic regression hyper‑parameters are tuned via 2‑fold cross‑validation over a log‑scale range of C. (5) Replicator Dynamics are iterated until the change in the strategy matrix falls below a tolerance ε or a maximum number of iterations K is reached. The final mixed strategies for the target nodes constitute soft predictions; the class with highest probability is taken as the hard label.

The authors evaluate the method—named GTDA—on two standard object‑recognition adaptation benchmarks: Office‑31 (Amazon, DSLR, Webcam domains) and Office‑Caltech (four domains, including Caltech256). Both shallow SURF descriptors (800‑dimensional) and deep ResNet‑50 features are used. Accuracy on the target domain is the evaluation metric. GTDA is compared against classic distribution‑alignment methods (CORAL, Subspace Alignment) and several deep adaptation approaches (DANN, RevGrad, JAN, DDC, etc.). Results show that GTDA without any prior already matches or slightly exceeds the baselines, while the +LR initialization consistently improves performance by 1–3 percentage points, often surpassing state‑of‑the‑art deep methods, especially when deep features are employed.

Key contributions of the work are: (i) introducing a principled game‑theoretic formulation for UDA that guarantees convergence to a Nash equilibrium, thereby ensuring label consistency; (ii) providing a parameter‑free pipeline that avoids intensive training phases typical of deep adaptation methods; (iii) demonstrating that soft pseudo‑labels from a simple classifier can effectively bootstrap the transductive game and boost accuracy; and (iv) achieving competitive results on widely used benchmarks with both shallow and deep feature representations. The paper also discusses limitations: constructing the full affinity matrix scales quadratically with the number of samples, which may hinder scalability to very large datasets, and the choice of local bandwidth and sparsification parameters, while automated, can still affect performance on heterogeneous data. Future directions include efficient graph approximations, multi‑source adaptation extensions, and integration of GTG outputs into subsequent supervised learning stages.