Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)

Aleksei Medvedev; Alejandro Ariza-Casabona; Steven Derby; Gonzalo Fiz Pontiveros; Xinyang Shao; Florian Spiess

cs.AIOct 1, 2026

Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)

Authors: Aleksei Medvedev, Alejandro Ariza-Casabona, Steven Derby, Gonzalo Fiz Pontiveros, Xinyang Shao, Florian Spiess

Organizations: Huawei Ireland Research Centre Dublin, Ireland

Abstract

Existing work on Semantic IDs (SIDs) for generative recommendation treats SID construction as a representation learning problem: encode items into a quantised latent space and read off codes. We argue this view is incidental. SID construction is, at heart, a recursive clustering problem, and once stated this way the natural object to cluster is a graph whose nodes carry semantic content and whose edges carry collaborative signal; SID assignment becomes a hierarchical graph partition. This reframing yields a unified framework, Graph-Informed Semantic IDs (GrIS), that subsumes prior approaches rather than displacing them. RQ-VAE and RQ-KMeans are recovered as the special case where the graph is empty, exposing content-only quantisation as one corner of a larger design space along two so-far-collapsed axes: graph construction and recursive partition algorithm. We explore two contrasting instantiations: RecDMoN, which performs hierarchical assignment via differentiable graph pooling, and RQ-GAE, which extends RQ-VAE with graph-aware item representations and a graph reconstruction objective. On multiple real-world datasets, GrIS consistently improves over CF-aware SOTA, with gains of up to +52% Hit@10. Because graph construction and partition are explicit, separately configurable components, improvements on either axis can be combined and evaluated systematically.

Figures & tables

Figure 1. Overview of the GrIS framework with two main axes: Graph Construction and Hierarchical SID Assignment. Graph Construction uses the user behaviour data to construct a weighted item graph based on predefined construction rules. This graph together with semantic item representations are fed to the second module to perform hierarchical partition. The output hierarchical SIDs are used to train a Generative Recommender model. In particular, we explore RQ-GAE and RecDMoN instantiations of the GrIS framework.

Method	Graph Construction	Hierarchical Partition
TIGER ( Rajput et al., 2023 )	$\emptyset$	RQ-VAE
LETTER ( Wang et al., 2024a )	$\emptyset$	RQ-VAE + Contrastive CF + Diversity Loss
MMGRec ( Liu et al., 2026 )	Multimodal Bipartite $\mathcal{U}\times\mathcal{I}$	RQ-VAE + BPR loss
S 2 GR ( Guo et al., 2026 )	Windowed Co-occurrence $\mathcal{I}\times\mathcal{I}$	RQ-VAE + Load Balancing Loss
RecDMoN (ours)	Immediate Transition $\mathcal{I}\times\mathcal{I}$	Differentiable Graph Pooling (DMoN)
RQ-GAE (ours)	Immediate Transition $\mathcal{I}\times\mathcal{I}$	RQ-VAE + Graph Reconstruction Loss

Table 1. Instantiations of the GrIS framework. Content-only methods arise as special cases with an empty graph (

E=\emptyset

Dataset	$\|\mathcal{U}\|$	$\|\mathcal{I}\|$	#Interactions	Avg. $T_{u}$	Sparsity
Toys	19,412	11,924	167,597	8.63	99.928%
Beauty	22,363	12,101	198,502	8.88	99.927%
Sports	35,598	18,357	296,337	8.32	99.955%
MIND	86,913	20,283	2,308,143	26.56	99.869%
Yelp	213,170	94,304	3,277,932	15.38	99.984%
Books	603,668	367,982	8,898,041	14.74	99.996%

Table 2. Dataset statistics.

Method	H@5	H@10	N@5	N@10	H@5	H@10	N@5	N@10
	Books				Beauty
RecDMoN	–	–	–	–	4.89 †	7.96 †	3.11 †	4.09 †
$\Delta_{\mathrm{LETTER}}$	–	–	–	–	+21.7%	+28.9%	+14.6%	+20.2%
RQ-GAE	3.31 †	4.60 †	2.39 †	2.81 †	4.43 †	6.87 †	2.96 †	3.74 †
$\Delta_{\mathrm{LETTER}}$	+29.0%	+34.1%	+26.1%	+29.0%	+10.2%	+11.2%	+9.1%	+9.9%
Spectral	–	–	–	–	4.17	6.35	2.76	3.45

Table 3. Results across datasets. Bold scores indicate leading results while underscore indicates second best results. † denotes significant improvement over the best baseline.

Figure 2. N@10 (%) on Beauty across graph construction strategies for (a) Recursive DMoN with different hierarchy configurations and (b) RQ-GAE with different APPNP propagation coefficients.

Graph variant	$\|E\|$	Avg. deg.	Avg. w-deg.	$\lambda_{2}$
adjacent	111649	18.45	21.72	0.0467
w2	198768	32.85	39.74	0.0588
w3	265133	43.82	54.07	0.0514
w5	361874	59.81	75.88	0.0536
w5_invdecay	361874	59.81	40.46	0.0537

Table 4. Graph statistics for the Beauty graph construction ablation.

Graph Input	Graph Loss	Beauty		Sports		Toys
Graph Input	Graph Loss	H@10	N@10	H@10	N@10	H@10	N@10
✓	✓	6.87	3.74	1.92	3.75	3.50	6.51
	✓	6.82	3.66	1.95	3.79	3.30	6.22
✓		5.66	3.10	1.49	2.95	2.78	5.38
		5.84	3.13	1.76	3.46	2.68	5.24

Table 5. RQ-GAE components ablation. Graph Input refers to a non-parametric graph-informed item representation (APPNP) and Graph Loss refers to the graph-contrastive loss at each codebook level. Bold scores indicate leading results while underscore indicates second best results.

Dataset	Method	$L_{1}$	$L_{2}$	$L_{3}$	$L_{4}$	$L_{5}$	Total
Beauty	Spectral	6	36	206	865	863	1976
	Gini	0.736	0.815	0.845	0.811	0.774
	RecDMoN	6	35	210	1245	33	1529
	Gini	0.099	0.159	0.194	0.269	0.632
Sports	Spectral	6	36	210	1087	661	2000
	Gini	0.517	0.655	0.730	0.727	0.813

Table 6. Per-level number of clusters and Gini index for Spectral Clustering and RecDMoN semantic ID mappings.

Method	Embedder	Beauty		Sports		Toys
Method	Embedder	H@10	N@10	H@10	N@10	H@10	N@10
RecDMoN	Qwen	7.96	4.09	4.72	2.41	8.16	4.02
RecDMoN	T5	7.24	3.71	4.47	2.31	7.29	3.65
RQ-GAE	Qwen	6.87	3.74	3.75	1.92	6.51	3.50
RQ-GAE	T5	6.85	3.72	3.57	1.85	6.51	3.43

Table 7. Results for RecDMoN and RQ-GAE with sentence-t5-base (T5) and Qwen3-Embedding-0.6B (Qwen) sentence embedders. Bold scores indicate leading results.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

	Model	L	Collision Rate	$\|\mathcal{C}^{l}\|$			Gini			Utilisation Ratio			$\delta_{Sem}$			$\delta_{CF}$
	Model	L	Collision Rate	$l=0$	$l=1$	$l=2$	$l=0$	$l=1$	$l=2$	$l=0$	$l=1$	$l=2$	$l<1$	$l<2$	$l<3$	$l<1$	$l<2$	$l<3$
Beauty	RecDMoN	5	0	7	36	211	0.224	0.179	0.194	-	-	-	0.074	0.122	0.177	-0.002	-0.006	0.005
	RQ-GAE	4	0	256	256	256	0.147	0.222	0.156	1	1	1	0.088	0.184	0.322	-0.010	0.001	-0.022
	LETTER	4	0	44	256	256	0.504	0.178	0.192	0.172	1	1	0.050	0.118	0.220	-0.005	0.012	0.018
	MMGRec	5	0	256	78	136	0.224	0.918	0.860	1	0.305	0.531	0.074	0.085	0.100	0.010	-0.002	0.001
	S 2 GR	4	0.001	44	250	244	0.297	0.197	0.187	0.172	0.977	0.953	0.163	0.333	0.481	0	-0.007	-0.001

Appendix

Table 8. SID-space diagnostics and per-level codebook cardinalities on all six datasets.

L

denotes the number of codebooks, while

\delta_{\mathrm{CF}}

and

\delta_{\mathrm{Sem}}

denote prefix-level cosine-similarity gaps computed using SASRec and Qwen3-Embedding-0.6B embeddings, respectively.

Explore similar work

Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation
Jul 27, 2026Junting Wang, Xinrui He, Yunzhe Li +1Sequential Recommendation Cross-Encoder Reranking
University of Illinois Urbana-Champaign Urbana, Illinois, USA
TopoGR: Revealing and Preserving Latent Structure of Semantic ID in Generative Recommendation
Jul 28, 2026Ziyu Zheng, Zhengshun Du, Yaming Yang +5Real-World Content Recommendation Problem Semantic Relationships
Xidian University, Xi’an, China · Alibaba Beijing, China · University of Science and Technology of China Hefei, China
FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation
Sep 29, 2026Song-Li Wu, Weinan Gan, Zhaocheng Du +2Codebooks Prior Knowledge Integration
Tsinghua University · Huawei Noah’s Ark Lab · University of Science and Technology of China

Jul 27, 2026, cs.AIJ/K move · Enter open · S save

Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

Junting Wang, Xinrui He, Yunzhe Li, Hari Sundaram

University of Illinois Urbana-Champaign Urbana, Illinois, USA

Sequential Recommendation Cross-Encoder Reranking Decoding Intrusion Detection

Open paper AskPDF arXiv

Semantic IDs (SIDs) are now a central component of generative recommendation. Current SID-based systems assign three roles to the same token sequence. Shared prefixes are intended to organize related items, the complete SID identifies an individual item, and each generated token narrows the items that can still be returned. We systematically investigate SIDs from item encoding and SID construction to autoregressive generation and final recommendation. We examine how SID construction changes item representations and how those changes affect generation. Across three Amazon domains and eight SID constructions, SID neighborhoods recover only 32.2% of the encoder's ten nearest neighbors on average. Alternative item descriptions still retrieve the corresponding item first in 99.57% of controlled cases, yet change 38.4% of exact SIDs. These results show that SIDs retain broad organization but lose much of the encoder's fine local structure, while their exact tokens are not determined by item meaning alone. This loss becomes consequential during generation. After the final semantic token, TIGER retains only 29.9% of held-out targets that were plausible recommendations before SID filtering. Motivated by these findings, we propose Item-Supported Decoding (ISD), a lightweight inference-time method that allows a user-specific item ranking to support corresponding SID prefixes before beam search discards them. The same ranking then orders the generated items. ISD requires no additional parameters or retraining of the SID constructor or decoder. We empirically show that ISD improves NDCG@10 over the corresponding SID backbone in every evaluated setting, with relative gains of up to 31.2%. Our results show that SIDs provide useful coarse item organization, but their fine boundaries should not alone determine which items remain available during generation.

Abstract

Explore similar work

Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

TopoGR: Revealing and Preserving Latent Structure of Semantic ID in Generative Recommendation

FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation