Can someone explain the difference between transductive learning and semi-supervised learning? Or is semi-supervised learning a type of transductive learning? Transductive learning is when we do not try to learn anything general enough but try to find labels of the unlabeled data. And semi-supervised is when there is small labeled data, a copious amount of unlabeled data, and we try to find labels of the latter using the former. The outcome (and the starting point) of both transductive and semi-supervised learning seem to be similar. What is the difference in the solution construction recipe (if any)? Latter focuses on empirical risk minimization by finding the optimal function that can explain the labels of labeled data and distribution of all the observed samples? What is the strategy of the former if it wants to escape learning anything general? Heuristics?