Recommender Systems with Generative Retrieval

TL;DR

The paper reframes sequential recommendation as a generative retrieval problem by representing items with Semantic IDs—quantized, content-informed token tuples generated via RQ-VAE. It introduces TIGER, a Transformer-based seq2seq model that predicts the next item's Semantic ID from a user's history, eliminating the need for explicit item indexing. Empirical results on three Amazon datasets show state-of-the-art retrieval performance and demonstrate benefits in cold-start scenarios and diversity control through decoding temperature. The work highlights the efficiency and generalization advantages of semantic, token-based item representations and points to avenues for further optimization and extensions.

Abstract

Modern recommender systems perform large-scale retrieval by first embedding queries and item candidates in the same unified space, followed by approximate nearest neighbor search to select top candidates given a query embedding. In this paper, we propose a novel generative retrieval approach, where the retrieval model autoregressively decodes the identifiers of the target candidates. To that end, we create semantically meaningful tuple of codewords to serve as a Semantic ID for each item. Given Semantic IDs for items in a user session, a Transformer-based sequence-to-sequence model is trained to predict the Semantic ID of the next item that the user will interact with. To the best of our knowledge, this is the first Semantic ID-based generative model for recommendation tasks. We show that recommender systems trained with the proposed paradigm significantly outperform the current SOTA models on various datasets. In addition, we show that incorporating Semantic IDs into the sequence-to-sequence model enhances its ability to generalize, as evidenced by the improved retrieval performance observed for items with no prior interaction history.

Figures (6)

• Figure 1: Overview of the Transformer Index for GEnerative Recommenders (TIGER) framework. With TIGER, sequential recommendation is expressed as a generative retrieval task by representing each item as a tuple of discrete semantic tokens.
• Figure 2: An overview of the modeling approach used in TIGER.
• Figure 3: RQ-VAE: In the figure, the vector output by the DNN Encoder, say ${\bm{r}}_0$ (represented by the blue bar), is fed to the quantizer, which works iteratively. First, the closest vector to ${\bm{r}}_0$ is found in the first level codebook. Let this closest vector be ${\bm{e}}_{c_0}$ (represented by the red bar). Then, the residual error is computed as ${\bm{r}}_1:={\bm{r}}_0-{\bm{e}}_{c_0}$. This is fed into the second level of the quantizer, and the process is repeated: The closest vector to ${\bm{r}}_1$ is found in the second level, say ${\bm{e}}_{c_1}$ (represented by the green bar), and then the second level residual error is computed as ${\bm{r}}_2={\bm{r}}_1-{\bm{e}}_{c_1}'$. Then, the process is repeated for a third time on ${\bm{r}}_2$. The semantic codes are computed as the indices of ${\bm{e}}_{c_0}, {\bm{e}}_{c_1},$ and ${\bm{e}}_{c_2}$ in their respective codebooks. In the example shown in the figure, this results in the code $(7,1,4)$.
• Figure 4: Qualitative study of RQ-VAE Semantic IDs $(c_1, c_2, c_3, c_4)$ on the Amazon Beauty dataset. We show that the ground-truth categories are distributed across different Semantic tokens. Moreover, the RQVAE semantic IDs form a hierarchy of items, where the first semantic token ($c_1$) corresponds to coarse-level category, while second/third semantic token ($c_2$/$c_3$) correspond to fine-grained categories.
• Figure 5: Performance in the cold-start retrieval setting.