View Generation in multi-view learning refers to techniques for creating additional views of data when natural multiple views are unavailable, artificially constructing diverse representations from a single data source to enable multi-view learning methods. View generation is essential because many multi-view algorithms (co-training, contrastive learning, CCA) require multiple views, but real-world datasets often come with only a single representation.

Why View Generation Matters in AI/ML: View generation enables multi-view learning when natural views don't exist, expanding the applicability of powerful multi-view methods (including modern contrastive learning) to single-view datasets through data augmentation, feature splitting, and learned transformations that create complementary representations.

• Data augmentation as views — The dominant approach in modern self-supervised learning: different random augmentations (cropping, color jittering, rotation, noise addition) of the same input create two "views" that share semantic content but differ in low-level details; SimCLR, BYOL, and MoCo all use this approach • Feature splitting — Dividing the feature set into disjoint subsets creates artificial views: e.g., splitting text features into word n-grams vs. character n-grams, or splitting tabular features into correlated groups; satisfies co-training's conditional independence assumption approximately • Random subspace views — Randomly projecting features into different low-dimensional subspaces creates diverse views; each projection captures different feature combinations, providing complementary perspectives similar to random forests' feature bagging • Learned view generators — Neural networks can learn to generate informative views: encoders trained with view-diversity objectives produce representations that are sufficiently different to provide complementary information while being sufficiently similar to agree on labels • Cross-modal generation — Generating missing modalities from available ones (text from images, depth from RGB) creates synthetic multi-modal views; this is increasingly practical with powerful generative models and enables multi-view learning on naturally single-modal data

View generation transforms single-view datasets into multi-view learning problems through data augmentation, feature splitting, and learned transformations, enabling the full power of multi-view methods—from classical co-training to modern contrastive self-supervised learning—on datasets that naturally provide only a single representation of each example.