PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.01533v1cs.AIcs.IRcs.LG

TL;DR

Semantic IDs are commonly built by quantising item representations, but this framing does not capture their role in recursively organising items for recommendation. GrIS treats SID construction as hierarchical partitioning of a graph combining semantic content and collaborative signal. Across multiple real-world datasets, graph-informed SIDs improve over strong collaborative-filtering-aware baselines, with gains up to 52% Hit@10.

  • Problem

    SID construction is commonly framed as latent-vector quantisation, although useful IDs must hierarchically organise items and account for collaborative signal.

    Source: §1 · pp. 1–2

  • Method

    GrIS formulates SID construction as hierarchical partitioning of an item graph and makes graph construction and recursive partitioning explicit design choices.

    Source: §1 · p. 2

  • Results

    GrIS improves over strong collaborative-filtering-aware baselines across multiple real-world datasets, with gains up to 52% Hit@10.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Content-only quantisation is an empty-graph special case, while graph construction and partitioning can be configured as separate design choices.

    Source: §1 · p. 2

    Collaborative graphs can become stale as behaviour changes, and the current RecDMoN implementation limits applicability to large item catalogues.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 Introduction

The paper reframes Semantic ID construction as recursive clustering over item graphs that combine semantic content with collaborative signal. GrIS unifies content-only methods and graph-informed approaches, with two distinct instantiations and consistent benchmark gains.

  • Existing SID construction treats items as latent vectors to quantise, but hierarchical codes must also preserve distinctions relevant to recommendation.Source: §1 · p. 1
  • GrIS formulates SID construction as hierarchical graph partition, with RQ-VAE and RQ-KMeans as empty-graph, content-only special cases.Source: §1 · p. 2
  • Separating graph construction from recursive partitioning makes these design choices independently configurable and comparable.Source: §1 · p. 2
  • RecDMoN assigns SIDs through recursive differentiable graph pooling, while RQ-GAE adds graph-aware representations and graph reconstruction to residual quantisation.Source: §1 · p. 2
  • Up to 52% Hit@10 improvement is reported for graph-informed SID construction over strong collaborative and SID-based baselines across multiple real-world datasets.Source: §1 · p. 2

2 Related Work

Prior SID work varies in whether it uses item content, behavioural signals, or graph structure. GrIS connects these approaches through hierarchical graph partition, while graph convolution and pooling provide operations for incorporating relations and forming clusters.

2 Related Work

  • The related-work review groups SID methods into content-based, behaviour-enhanced, and structure-aware families, alongside graph convolution and pooling operations.Source: §2 · p. 2

2.1 Content-based Semantic IDs

  • Content-based SIDs quantise item content into discrete codes, but their semantic similarity may not align with user behaviour.Source: §2.1 · p. 2

2.2 Behaviour-enhanced Semantic IDs

  • Behaviour-enhanced methods align semantic quantisation with collaborative signals, including CF embeddings in LETTER and co-occurrence contrastive regularisation in PLUM.Source: §2.2 · p. 2

2.3 Structure-aware Semantic IDs

  • Structure-aware methods use interaction topology in ID construction; GrIS unifies these lines as hierarchical graph partition, with content-only methods as empty-graph cases.Source: §2.3 · pp. 2–3

2.4 Graph Convolution and Graph Pooling

  • Graph convolution updates node representations using neighboring features, while graph pooling learns soft cluster assignments and coarsens graph connectivity.Source: §2.4 · p. 3

3 Preliminaries

The paper studies SID construction for sequential recommendation, where the next item is predicted from a user’s interaction history. Its formulation combines item content embeddings with a configurable collaborative graph and evaluates SID quality while holding the downstream recommender fixed.

  • Sequential recommendation predicts a user’s next item from the prefix of that user’s interaction history.Source: §3 · p. 3
  • Items have semantic embeddings derived from content; this work uses text embeddings from dataset-dependent metadata fields.Source: §3 · p. 3
  • The collaborative graph represents behaviourally related items with weighted edges, and its construction rule is configurable.Source: §3 · p. 3
  • The objective is a hierarchical SID map that improves retrieval quality, with the downstream generative recommender held fixed.Source: §3 · p. 3

4 Graph Informed Semantic IDs (GrIS)

GrIS frames Semantic ID construction as hierarchical graph partition, with graph construction and recursive partition as independent design choices. It covers content-only methods as the empty-graph case and includes both residual-quantisation and graph-pooling instantiations.

4.1 Overview

  • GrIS separates graph construction from hierarchical partition and treats content-only methods as the empty-graph case.The two axes determine what relational structure is encoded and how it is divided into codes.Source: §4.1 · p. 3, Table 1 · p. 4

4.2 Item Graph Construction

  • User interaction sequences define weighted item-item edges under a chosen rule, with low-weight edges optionally removed by a threshold.Rules include windowed co-occurrence, sequential transitions, session co-appearance, and random-walk proximity.Source: §4.2 · p. 3

4.3 Hierarchical SID assignment

  • Given a graph and initial item representations, GrIS assigns hierarchical SIDs that are used to train a generative recommender.Variable-length and end-to-end SID construction are left for future integration.Source: §4.3 · p. 3, Figure 1 · p. 4

4.3.1 Graph-Informed Representations.

  • Graph-informed representations propagate information across neighboring items to enrich initial embeddings with local collaborative context.Propagation may use a parametric encoder or normalized neighborhood aggregation with fixed weights.Source: §4.3.1 · p. 4

4.3.2 Residual Quantisation Instantiation.

  • RQ-GAE combines residual quantisation with graph reconstruction supervision, using a subsampled neighbourhood softmax over batch items rather than reconstructing the full adjacency matrix.The graph term encourages the quantised latent space to preserve the item-item interaction graph; β controls its strength.Source: §4.3.2 · pp. 4–5

4.3.3 Recursive DMoN Instantiation.

  • Recursive DMoN repeatedly clusters each parent graph, assigns items to child clusters, and forms each SID from the item’s root-to-leaf branch sequence.Its local objective maximizes relaxed graph modularity while penalizing highly imbalanced soft assignments; shared prefixes indicate items that remained in the same partition.Source: §4.3.3 · p. 5

5 Experiments

Experiments across recommendation benchmarks compare graph-informed SID methods with semantic, collaborative, and graph-aware baselines. Results vary by dataset and method, while graph construction, partition design, and scale shape performance.

5.1 Experimental Setup

  • Benchmarks include four Amazon Reviews 2014 subsets (Beauty, Toys, Sports, Books), MIND-Small, and Yelp; after 5-core filtering, leave-one-out evaluation tests each user’s last interaction and reports H and N at cutoffs 5 and 10.Source: §5.1 · pp. 5–6

5.2.1 Generative Recommendation Comparison.

  • 4.60 H@10 on Books makes RQ-GAE strongest there, with +16.5% over S2GR and +34.1% over LETTER; RecDMoN leads on Beauty, Sports, and Toys.RQ-GAE ranks among the top two on Books, Beauty, Sports, and Yelp; HSTU is strongest overall on MIND while RecDMoN remains competitive.Source: §5.2.1 · p. 6, Table 3 · p. 7

5.2.2 Impact of Graph Construction and Graph-Aware Design Choices.

  • Across Beauty, Toys, and Sports, the full RQ-GAE model ranks first on Beauty and Toys and remains competitive on Sports; graph loss alone consistently improves over the base model.Graph-informed inputs alone yield mixed results, underperforming the base model on Beauty and Sports, while graph input and graph loss together contribute complementary structural information and graph-proximity regularisation.Source: Table 5 · p. 8, §5.2.3 · p. 8

5.2.4 Graph Clustering Analysis.

  • RecDMoN produces substantially lower Gini indices than spectral clustering from L1 to L4 across Beauty, Sports, and Toys, indicating more evenly distributed clusters.RecDMoN allocates more clusters to L4 and fewer to the L5 deduplication codebook, while both methods show higher inequality at the finest level.Source: Table 6 · p. 8, §5.2.4 · p. 8

5.2.5 Sentence Embedding Ablation.

  • RecDMoN performs best on smaller datasets but its dense adjacency matrix makes it impractical for larger catalogs, where RQ-GAE is competitive or dominant.The methods are distinct partitioning families: RecDMoN uses global differentiable soft assignments, whereas RQ-GAE retains greedy residual quantisation and adds graph-aware reconstruction.Source: §5.2.6 · pp. 8–9

5.2.2 Impact of Graph Construction and Graph-Aware Design Choices.

  • On Beauty, deeper RecDMoN L4_RF6 performs strongest, while RQ-GAE generally performs better with smaller APPNP propagation coefficients.Very large coefficients make representations closer to the initial semantic embeddings and are associated with less stable downstream performance; graph statistics show no simple monotonic relation to recommendation quality.Source: §5.2.2 · pp. 7–8, Figure 2 · p. 9

5.2.5 Sentence Embedding Ablation.

  • Replacing sentence-t5-base with Qwen3-Embedding-0.6B consistently improves both RecDMoN and RQ-GAE across Beauty, Sports, and Toys.Gains are especially pronounced for RecDMoN; RQ-GAE has smaller but consistent improvements.Source: §5.2.5 · p. 8, Table 7 · p. 9

6 Conclusions

GrIS frames semantic-ID construction as hierarchical partitioning of a collaborative item graph, with content-only methods as empty-graph cases and graph construction and partition as explicit design axes.

6 Conclusions

  • GrIS formalises semantic-ID construction as hierarchical graph partition and treats content-only approaches as empty-graph special cases.Its two explicit design axes are graph construction and hierarchical partition.Source: §6 · p. 9
  • Future graph work could replace co-occurrence edges with knowledge graphs and examine longer-range session co-appearance or random-walk relations.Source: §6 · p. 9
  • Partitioning item n-grams into k codes could shorten sequences by a factor of n, providing a direct lever on transformer computational cost.The passage gives the sequence-length factor as k/n rather than k and notes transformer cost scales as Ω(L^2).Source: §6 · p. 9
  • Collaborative methods face temporal drift as changing behaviour makes item graphs and derived partitions stale, while current RecDMoN limits applicability to large catalogues.Incremental graph and partition updates and sparse, sampled, or mini-batch RecDMoN variants are left for future work.Source: §6 · p. 9

7 GenAI Usage Disclosure

The authors report using generative AI for language polishing and routine code assistance, while retaining responsibility for the research and its conclusions.

7 GenAI Usage Disclosure

  • Generative AI assisted with language polishing, grammar, writing suggestions, and simple routine code; the authors reviewed and verified the outputs and take responsibility for the research.Source: §7 · p. 9

A SID Quality Metrics

The appendix assesses codebook use, SID collisions, and prefix-level collaborative and semantic coherence, reporting differences in the balance and comparability of methods.

A SID Quality Metrics

  • Table 8 evaluates codebook cardinality, utilisation, usage Gini, and complete-SID collision rate across six datasets.Higher utilisation and lower Gini indicate better coverage and balance; lower collision rate is better.Source: A SID Quality Metrics · p. 11
  • Prefix coherence is δ = μ_intra − μ_inter, estimated from 1,000 sampled item pairs using SASRec embeddings for δCF and Qwen3-Embedding-0.6B for δSem.A larger positive δ indicates greater similarity among items sharing a prefix than among items assigned to different prefixes; the metrics complement downstream accuracy.Source: A SID Quality Metrics · p. 11
  • Graph-aware methods preserve more semantic coherence, and RQ-GAE has the strongest overall CF–semantic balance while maintaining balanced codeword usage.The passage states this balance could explain RQ-GAE's good overall downstream performance.Source: A SID Quality Metrics · p. 11
  • RecDMoN and spectral clustering use fewer than 10 first-level codewords, making their δ values less comparable with finer-grained methods; RecDMoN has higher δSem and δCF between the two.Source: A SID Quality Metrics · p. 11

Related papers

Loading 2610.01533v1…