Semantic ID
A representation in generative recommendation that maps discrete items into token sequences, typically produced by quantization methods such as RQ-VAE, residual quantization, or hierarchical clustering. Each codebook level corresponds to a granularity level, with earlier tokens encoding coarse categories and later tokens encoding fine-grained differences, turning item prediction into a sequence generation problem.
Timeline (last 90 days)
Most existing generative recommendation methods only use the standard causal cross-entropy objective to optimize semantic ID prediction, ignoring the hierarchical structure of the semantic IDs themselves.
Standard causal cross-entropy treats the token at every position in the sequence equally, without distinguishing which level of the semantic ID hierarchy a token is at, flattening a prediction task that could be decomposed into a tree structure into mutually independent classification problems.
The quality of semantic IDs directly determines the upper bound of generative recommendation, because they are both the input representation and the generation target.
The core idea of generative recommendation is to encode items into semantic IDs and then generate these IDs layer by layer in an autoregressive manner.
All Facts (4)
Most existing generative recommendation methods only use the standard causal cross-entropy objective to optimize semantic ID prediction, ignoring the hierarchical structure of the semantic IDs themselves.
50%UnverifiedStandard causal cross-entropy treats the token at every position in the sequence equally, without distinguishing which level of the semantic ID hierarchy a token is at, flattening a prediction task that could be decomposed into a tree structure into mutually independent classification problems.
50%UnverifiedThe quality of semantic IDs directly determines the upper bound of generative recommendation, because they are both the input representation and the generation target.
50%UnverifiedThe core idea of generative recommendation is to encode items into semantic IDs and then generate these IDs layer by layer in an autoregressive manner.
50%