Can semantic IDs enhance knowledge graph paths for generative recommendation?

SPRIG swaps opaque item tokens for content-derived codes

Generative recommenders predict the next item directly as a token sequence rather than scoring a catalog. Two research lines have tried to make that prediction smarter. Semantic IDs (SIDs) replace arbitrary, randomly initialized item embeddings with hierarchically quantized discrete codes derived from item content, so similar items share code prefixes. Knowledge-graph (KG) path reasoning instead generates entity-relation paths that ground recommendations in structured relationships between items, attributes, and external entities.

Each line has a complementary weakness. SID-based models lack relational grounding, while KG-based generative recommenders still represent items as opaque tokens tied to large embedding tables, which limits parameter sharing and generalization. SPRIG integrates the two: it rewrites every item entity in a KG path as its SID sequence, so a path like <user, watched, Inception> becomes <user, watched, SEM1, SEM2, SEM3>. The model is a causal Transformer decoder trained on these SID-augmented paths, first pretrained on generic KG paths without collaborative signal, then fine-tuned on user-to-item paths. At inference it emits a SID sequence that an inverse lookup maps back to a catalog item.

The item vocabulary stops growing with the catalog

In KGGLM, the primary KG baseline, each item gets its own embedding, so the item-dependent vocabulary grows linearly with the catalog: 2,984 entries on ML-1M and 18,415 on Onion. SPRIG instead represents every item with Ls = 3 codes drawn from K = 256 shared entries, bounding the item-token vocabulary by Ls · K = 768 regardless of catalog size.

That is roughly a 4x reduction on ML-1M (2,984 to 768) and roughly 24x on the larger Onion catalog (18,415 to 768). The paper reports this compression does not cost capacity: on ML-1M SPRIG is comparable in size to KGGLM (56.1M vs. 57.8M parameters) yet uses 11% fewer FLOPs per step while reaching higher accuracy. On Onion, where KGGLM's per-item embeddings dominate its parameter count, SPRIG uses 19% fewer parameters and nearly half the FLOPs. The efficiency gap widens as the catalog grows.

SID item tokens raise nDCG@10 by 44% and 58%

The paper's first research question asks whether replacing opaque item tokens in KG-based generative recommenders with semantically grounded IDs improves recommendation quality. On ML-1M, the fully specialized SPRIG (gen + spec) improves nDCG@10 by roughly 44% over its directly comparable baseline KGGLM (gen + spec). On the larger and sparser Onion dataset, SPRIG raises nDCG@10 by 58%. The paper states these improvements over KGGLM (gen + spec) are statistically significant across all metrics on both datasets.

Because KGGLM and SPRIG share the same backbone and knowledge-graph signal, the paper attributes the gain to the SID-based item representation and its associated decoding rather than to the knowledge graph. Even without collaborative fine-tuning, SPRIG (gen-only) surpasses KGGLM (gen-only) on every metric on both datasets; for example, nDCG@10 improves from 0.0281 to 0.0375 on ML-1M. The paper attributes this to parameter sharing induced by SIDs, which lets rarely seen items inherit signal from prefix-level neighbors.

Relational grounding adds signal that SID-only models miss

Across representation paradigms, the generative baselines OpenP5 (atomic-ID) and TIGER (SID-based) are substantially outperformed by the knowledge-aware models, which the paper reads as confirmation that relational grounding adds signal that SID-based sequential models alone miss. Among directly comparable generative models, the gap between SPRIG and KGGLM shows that relational grounding and SIDs are complementary and that combining them is strongest.

SASRec attains the highest nDCG@10 on both datasets and the highest HR@10 on Onion, but the paper notes it ranks the full catalog directly while the generative models decode over beam-search candidates, so that comparison is indicative rather than exact. The paper further notes SASRec's margin on Onion reflects the sequentiality of music listening that the sampled setup discards.

SPRIG tolerates quantizer choice but not weak item features

To test robustness to the quantization approach, the authors evaluate three residual quantizers for semantic-ID construction: RVQ, RQ-VAE, and residual K-means. Consistent with GRID, residual K-means attains the best accuracy, but the gap over RVQ and RQ-VAE is small (about 1%), indicating SPRIG is largely insensitive to the choice of quantizer.

Robustness to item representation behaves differently. Comparing two audio feature extractors on Onion, musicnn embeddings consistently outperform the MFCC bag-of-words representation across all quantizers. The paper reads this as SPRIG being sensitive to the quality of the underlying item features, so the quantizer is a low-stakes choice while the content encoder is not.

Cold start and readable paths remain untested

The evaluation covers two domains: movies on ML-1M (6,040 users, 2,984 items, 932,295 interactions) and music on Onion (12,784 users, 18,415 items, 1,375,341 interactions). The paper's claims about parameter and compute savings are tied to these catalogs and to the Ls = 3, K = 256 SID configuration used throughout.

Two limits qualify the results. First, the paper presents item cold start as a natural and promising application of SIDs, since a new item can be reached through the SID prefix it inherits from content-similar seen items, but states that a faithful evaluation requires cold-start-specific splits and a path-sampling strategy for items absent from the user-item training edges, which is left to future work. Second, although KG path reasoning is inherently explanatory, SPRIG's paths terminate in SID codes that are less human-readable than named item entities; recovering faithful, interpretable justifications is described as a promising but non-trivial direction rather than an existing feature. The paper also notes that its optimal quantizer differs from that reported for purely sequential SID models, motivating further study of how tokenizer choice interacts with relational path modeling.

The vocabulary saving matters most on large catalogs

The decision-relevant pattern in this paper is a size-dependent tradeoff. On ML-1M, SPRIG's advantage over KGGLM is a modest parameter difference (56.1M vs. 57.8M) plus 11% fewer FLOPs per step and higher accuracy. On Onion, where per-item embeddings dominate KGGLM's parameter count, the same design yields 19% fewer parameters and nearly half the FLOPs, alongside a 58% nDCG@10 gain.

For readers weighing whether to adopt SID-augmented KG paths, the paper's evidence points to catalog size as the variable that decides how much the fixed 768-entry item vocabulary buys: the larger the catalog, the larger the parameter, compute, and accuracy return. The paper's own caveat is that this return depends on item feature quality, since musicnn embeddings consistently beat MFCC bag-of-words across all quantizers, while the choice among RVQ, RQ-VAE, and residual K-means moves accuracy by only about 1%.

Sources used in this answer