What is embedding-based classification architecture?

Embedding-based classification converts inputs into dense vector representations before classification.

Instead of using raw features, embedding-based architectures map inputs (text, users, items) into dense vectors in a latent space. These embeddings capture semantic similarity and are fed into downstream classifiers like MLPs or similarity-based models. Embeddings can be learned end-to-end or pre-trained (e.g., BERT, Word2Vec). This approach improves generalization and is widely used in NLP and recommendation systems.