The paper addresses the challenge of data scarcity in sports analytics by leveraging transfer-learning techniques. Particularly, we investigate the application of the promising Domain-Adversarial Neural Networks (DANN) technique to enhance the generalization capabilities of machine learning models across different but related sports, which is useful to transfer knowledge from a data-richer to a data-poorer domain. As a test-bench, we built two different-size datasets, one per sport type, with the one in the tennis domain containing twice as many sample as the one in table-tennis. We utilized a wearable device to collect sensor data for the same five types of shots in both tennis and table tennis. Experimental results demonstrate a 25% improvement in tennis-table shot classification accuracy for generalization to new players when using the DANN-enhanced model compared to the original one, effectively reducing the overfitting issues typically associated with small datasets. As key findings, we highlight the potential of DANN in extracting domain-invariant features that allow improving performance of activity classifiers also for sports with relatively poor datasets.