TIGER

Recommender Systems with Generative Retrieval

Shashank Rajput et al. · NeurIPS 2023

Read paper · arXiv:2305.05065

Encode items as residual-quantized semantic IDs, then generate IDs and retrieve valid catalog items.

What you will implement

Frozen residual encoding and collision suffixes, then unconstrained beam search and exact-ID filtering; no RQ-VAE or recommender training.

• 64. TIGER: Residual Quantization and Semantic IDs

  Quantize the remaining residual at each level and assign unique suffixes to colliding items.

  Frozen residual encoding from Section 3.1 and collision suffixes from Section 4; no encoder/decoder/codebook training.

• 65. TIGER: Generate IDs and Retrieve Valid Items

  Generate complete IDs with cumulative conditional log-probabilities, then filter invalid and duplicate items.

  Sections 3.2 and 4.5: unconstrained full-ID beam decoding followed by exact catalog filtering; fixed logits, no seq2seq training or prefix matching.