PaperlayerView exact revision on arXiv

Source-linked AI summary

ReliGRec: Reliability-Oriented LLM-Based Generative Recommendation via User-Risk-Aware Prompt Routing

Haoran Yang, Fei Chen, Yutian Xiao, Jiahao Liang

arXiv:2609.16560v1cs.IR

TL;DR

User-level weak-risk-aware generation control remains underexplored in LLM-based recommendation. ReliGRec estimates weak risk from behavioral and temporal collaborative evidence to route prompts before generation; the paper reports competitive recommendation and proxy-label prediction, with recommendation results varying across Beauty and Yelp.

  • Problem

    Pre-decoding generation policies based on user-level weak risk remain underexplored, despite histories that can contain unstable or inconsistent evidence.

    Source: Introduction · pp. 1–2

  • Method

    ReliGRec fuses Behavior and temporal Graph Tokens to estimate weak risk from partially labeled users and route Simple or Cautious Prompts before generation.

    Source: Introduction · p. 2, Problem Definition · p. 2

  • Results

    The paper reports competitive recommendation and proxy-label prediction; complete-system results favor ReliGRec across five metrics on Beauty and are comparable to the strongest baseline on Yelp.

    Source: Overall Performance (RQ1) · p. 6, Conclusion · p. 7

  • Takeaways & Limitations

    Weak risk serves as a generation-time prompt-routing signal, while the aggregated Graph Token supplies collaborative context for next-item generation.

    Source: Introduction · p. 2, Conclusion · p. 7

    The reported routing evidence does not establish recommendation improvement over the strongest uniform policy or matched-ratio random routing.

    Source: Prompt-Routing Analysis (RQ3) · p. 6

Abstract

from arXiv · show

User behavior in real-world recommender systems is heterogeneous. While some users exhibit coherent preferences, others show abrupt interest shifts, bursty interactions, excessive repetition, or inconsistency with collaborative neighborhoods. Such deviations may arise from benign variation or manipulation, including shilling attacks, but do not alone establish malicious intent. Existing robust recommenders exploit user-risk signals through training-time reweighting or graph aggregation, whereas adapting generation to estimated user-level weak risk remains underexplored in LLM-based generative recommendation. We propose ReliGRec (Reliability-oriented Generative Recommendation), a weakly supervised framework whose name denotes its design goal rather than a supervised reliability variable. ReliGRec derives user-level weak-risk proxy labels from review-feedback signals for a subset of users and represents sequential behavior and collaborative context using a Behavior Token and temporal Graph Tokens, respectively. A Dual-View Weak-Risk Estimator fuses the representations to produce a user-level weak-risk score that selects a Simple or Cautious Prompt at inference. The Cautious Prompt is designed to encourage attention to stable, collaboratively supported evidence while reducing overreliance on isolated, short-term, or repeated interactions. The Behavior Token affects generation through weak-risk estimation and routing, whereas the aggregated Graph Token provides collaborative context for next-item Semantic ID generation. ReliGRec thus turns weak-risk estimation from an auxiliary prediction into a generation-time control signal. Experiments report competitive recommendation and weak-risk proxy-label prediction, while routing analyses characterize the recommendation-quality and inference-cost behavior of weak-risk-guided prompting.

Introduction

LLM-based recommenders condition on histories that may contain unstable or inconsistent evidence, while adapting generation to user-level weak risk remains underexplored. ReliGRec estimates that risk from behavioral and temporal collaborative representations and uses it to route generation prompts.

Introduction

  • Histories can include abrupt shifts, bursty activity, repetition, or collaborative inconsistency, so uniform prompting may over-rely on unstable evidence.Such patterns may arise from benign variation or manipulation, including shilling attacks.Source: Introduction · p. 1, Figure 1 · p. 1
  • Adapting pre-decoding generation to user-level weak risk while retaining dynamic collaborative evidence remains underexplored.Prior robust recommenders mainly intervene during training, while output-side methods use generation uncertainty rather than user-level weak risk for routing.Source: Introduction · pp. 1–2
  • Weak risk captures relative deviation from regular interaction patterns using review-feedback proxy labels, not verified malicious status or calibrated attack probability.Behavioral cues alone may be ambiguous, motivating temporal collaborative context and partial weak supervision.Source: Introduction · p. 2
  • ReliGRec fuses a Behavior Token and temporal Graph Tokens in a Dual-View Weak-Risk Estimator to select a Simple or Cautious Prompt.The estimator is intended to address ambiguity in single-view evidence.Source: Introduction · p. 2
  • The Cautious Prompt is designed to reduce reliance on isolated, transient, or repeated evidence and prioritize temporally stable, collaborative support.The aggregated Graph Token also conditions next-item Semantic ID generation.Source: Introduction · p. 2
  • The work evaluates recommendation and weak-risk proxy-label prediction on Beauty and Yelp and characterizes routing quality–efficiency behavior.Its stated contribution distinguishes pre-decoding prompt routing from training-time risk weighting and output-side uncertainty handling.Source: Introduction · p. 2

Problem Definition

The problem is framed around predicting a user's next item from an observed interaction prefix and estimating user-level weak risk with partially labeled users.

Problem Definition

  • For a user, the task uses the observed interaction prefix up to position t to predict the next item.Items are represented by length-L Semantic IDs, following the item-identifier formulation of generative retrieval.Source: Problem Definition · p. 2
  • Weak-risk estimation uses review-feedback-derived labels for a subset of users, with labels denoting lower or higher proxy deviation from regular interaction patterns.These labels do not identify verified malicious users.Source: Problem Definition · p. 2
  • The remaining users are unlabeled, making the estimation setting partially labeled.The labeled set is denoted UL, with yu values 0 and 1 for the lower- and higher-proxy-deviation groups.Source: Problem Definition · p. 2

Methodology

ReliGRec uses behavioral and temporal collaborative representations to estimate user-level weak risk and condition next-item generation.

Methodology

  • The Behavior Token and temporal Graph Tokens jointly support user-level weak-risk estimation.Source: Methodology · p. 2
  • The estimated weak risk controls routing between the Simple and Cautious Prompts before generation.Source: Methodology · p. 2
  • The selected prompt and projected Graph Token condition next-item generation, while the Behavior Token acts through risk estimation and routing.Source: Methodology · p. 2

Semantic ID Construction

ReliGRec constructs hierarchical Semantic IDs with an offline item tokenizer and uses them to represent historical items and generation targets.

Semantic ID Construction

  • Raw item IDs are only indices and do not encode semantic or collaborative item relationships.Source: Semantic ID Construction · p. 2
  • An offline RQ-VAE tokenizer uses item semantic features and collaborative embeddings to construct Semantic IDs.The losses include reconstruction, residual quantization, and alignment with collaborative embeddings.Source: Semantic ID Construction · p. 2
  • RQ-VAE quantizes residuals into codebook indices, which form hierarchical Semantic IDs stored with associated items in a lookup table.The IDs represent both historical items and next-item generation targets.Source: Semantic ID Construction · p. 2
  • The framework uses hierarchical Semantic IDs, with the aggregated Graph Token providing collaborative information for next-item Semantic ID generation.Source: Figure 2 · p. 3

Dual-Path User Representation

ReliGRec represents user behavior and collaborative context in parallel using a Behavior Token and temporal Graph Tokens. The resulting representations provide complementary evidence for weak-risk estimation.

Temporal Graph Token Encoding

  • A shared graph encoder processes user–item graphs built from overlapping interaction windows to capture evolving collaborative relations.Each window's Graph Token retains the user's structural state and window-level item context.Source: Temporal Graph Token Encoding · p. 3
  • Temporal Graph Tokens summarize evolving collaborative structures, providing dynamic evidence for weak-risk estimation and next-item generation.Chronologically organized window representations are aggregated by a temporal Transformer, whose final-window output serves as the user representation.Source: Temporal Graph Token Encoding · p. 3

Behavior Token Encoding

  • The Behavior Token summarizes the user's recent ordered interactions as behavioral evidence complementary to the collaborative representation.A Transformer encodes the most recent items with item and positional embeddings, using the last valid hidden state.Source: Behavior Token Encoding · pp. 3–4

Weak-Risk Estimation and Weak-Risk-Guided Prompt Routing

ReliGRec fuses behavioral and temporal collaborative representations into a user-level weak-risk score, then uses that score to route each user to a prompt before decoding. Weak-risk labels supervise the estimator for labeled users, while unlabeled users remain available for recommendation training.

Dual-View Weak-Risk Estimator

  • Behavioral and temporal collaborative representations are projected into a shared space, concatenated, and passed through a two-layer MLP to produce a user-level weak-risk score.The score measures agreement with proxy-deviation patterns, not a calibrated attack probability or confirmed malicious judgment.Source: Dual-View Weak-Risk Estimator · p. 4
  • Labeled recommendation instances inherit their users' weak-risk labels and train the estimator with positive-class-weighted binary cross-entropy.Unlabeled users are excluded from the weak-risk loss but remain available for recommendation training.Source: Dual-View Weak-Risk Estimator · p. 4

Weak-Risk-Guided Prompt Routing

  • At inference, the predicted score selects the Cautious Prompt for users at or above the fixed threshold and the Simple Prompt otherwise.The operational cutoff is τ = 0.5, not a calibrated posterior threshold.Source: Weak-Risk-Guided Prompt Routing · p. 4
  • Training assigns prompts from weak-risk proxy labels, whereas inference assigns them from predicted scores.Because routing is hard, the recommendation loss does not optimize the weak-risk estimator through the routing decision.Source: Weak-Risk-Guided Prompt Routing · p. 4
  • The Cautious Prompt instructs the model to favor stable patterns and treat the Graph Token as supplementary context, excluding candidates supported mainly by short-term or noisy signals.Both prompt templates require one valid Semantic ID without intermediate analysis; the instructions express an intended evidence-use policy, not a verified reasoning process or malicious-identity judgment.Source: Weak-Risk-Guided Prompt Routing · p. 4
  • The Simple Prompt asks for recommendations directly from the available history and Graph Token.The Cautious Prompt instead marks input as potentially unreliable and asks the model to select an item aligned with longer-term interests.Source: Weak-Risk-Guided Prompt Routing · p. 4

Graph-Conditioned Semantic ID Generation

ReliGRec projects the aggregated Graph Token into the LLM's hidden space and supplies it alongside the selected prompt and historical Semantic IDs for next-item generation.

Graph-Conditioned Semantic ID Generation

  • A two-layer projector maps the aggregated Graph Token into the LLM hidden space, where it replaces a reserved Graph Token placeholder.The selected prompt, historical-item Semantic IDs, and projected Graph Token jointly condition an instruction-tuned LLM that autoregressively predicts the next item's discrete codes.Source: Graph-Conditioned Semantic ID Generation · p. 4

Training Objectives and Constrained Inference

ReliGRec trains its Semantic ID tokenizer offline and uses autoregressive generation with constrained decoding to produce next-item recommendations. Inference filters invalid prefixes and previously interacted items before returning recommendations.

Training Objectives

  • The Semantic ID tokenizer is trained offline, and the recommendation objective uses the target Semantic ID sequence with an end-of-sequence token during teacher forcing.The recommendation loss is mean cross-entropy over unmasked target tokens.Source: Training Objectives and Constrained… · p. 4

Evaluation Setup

  • Table 1 reports statistics for the processed datasets used in the evaluation.Source: Table 1 · p. 5

Training Objectives

Constrained Inference

  • At inference, a prefix Trie restricts decoding to valid Semantic ID prefixes, ranks completed sequences by summed token log-probabilities, maps them to items, and filters prior interactions.Source: Training Objectives and Constrained… · p. 5

Experiments

ReliGRec is evaluated on Beauty and Yelp for next-item recommendation and weak-risk proxy-label prediction, with controlled routing and graph-conditioning analyses on Beauty. It leads all five recommendation metrics on Beauty, performs comparably to the strongest baseline on Yelp, and its routing analysis does not establish gains over All-Simple or matched-ratio random routing.

Experimental Settings

  • Weak-risk proxy labels use review-vote ratios and majority evidence for users with at least three labeled reviews; remaining users are unlabeled and participate only in recommendation training.The labeled-user split is user-disjoint and uses an 8:2 training-to-test ratio.Source: Experimental Settings · p. 5, Table 2 · p. 5

Overall Performance (RQ1)

  • On Beauty, ReliGRec improves H@1, H@5, H@10, N@5, and N@10 over LETTER-TIGER by 54.67%, 19.24%, 8.93%, 27.28%, and 19.29%, respectively; on Yelp, it ties for best H@1 and ranks second on four metrics.The Yelp gaps from corresponding best results range from 0.12% to 1.87%; complete-system comparisons do not isolate routing, graph conditioning, or tokenizer and backbone differences.Source: Overall Performance (RQ1) · p. 6

Weak-Risk Estimation (RQ2)

  • ReliGRec achieves the best AUPRC, AUROC, and F1 on both user-disjoint weak-risk test sets; fused behavior and graph inputs yield the best Beauty scores across all three metrics.These metrics measure agreement with user-level weak-risk proxy labels, not verified fraud-detection performance.Source: Weak-Risk Estimation (RQ2) · p. 6

Prompt-Routing Analysis (RQ3)

  • Weak-risk-guided routing does not outperform All-Simple or matched-ratio random routing on NDCG@10, and the learned score's 66.9% Cautious-route rate is not a prevalence estimate.Its NDCG@10 is 0.0370 versus 0.0379 for All-Simple; end-to-end costs are 0.460, 0.548, and 0.479 seconds per user for All-Simple, All-Cautious, and weak-risk-guided routing, respectively.Source: Prompt-Routing Analysis (RQ3) · p. 6

Graph-Conditioning and Temporal-Aggregation Ablation (RQ4)

  • On Beauty, temporal graph conditioning improves all four recommendation metrics over a history-only generator, but the static Graph Token leads three ranking metrics while the temporal variant leads Hit@10.Temporal aggregation yields higher weak-risk AUPRC, AUROC, and F1 than the static variant in this run.Source: Graph-Conditioning and… · p. 6

Related Work

Prior generative-recommendation work develops Semantic IDs and collaborative conditioning, while risk-aware recommenders intervene mainly during training or use generation uncertainty during or after decoding. ReliGRec instead estimates user-level weak risk before decoding and uses it to route prompts.

LLM-based Generative Recommendation

Robust and Risk-Aware Recommendation

  • Robust recommenders use risk signals through joint detection, graph aggregation, reweighting, or representation rectification, mainly during training (Zhang et al. 2020; Qin et al. 2026).GUIDER and UGR use generation uncertainty for reranking, diversity control, rejection, or truncation during or after generation (Fan et al. 2026).Source: Robust and Risk-Aware Recommendation · p. 7
  • ReliGRec estimates user-level weak risk from sequential behavior and temporal collaborative context before decoding, then routes between Simple and Cautious Prompts while retaining Graph Token conditioning.This contrasts with training-stage risk interventions and uncertainty-based controls during or after generation.Source: Robust and Risk-Aware Recommendation · p. 7

Conclusion

ReliGRec combines user-level weak-risk estimation, collaborative modeling, and prompt routing before generation. Its results show competitive recommendation and proxy-label prediction, while routing does not establish an advantage over All-Simple or matched-ratio random routing.

Conclusion

  • ReliGRec combines temporal Graph Tokens for collaborative conditioning with Behavior and Graph representations for weak-risk estimation from partially labeled users.The weak-risk score selects a Simple or Cautious Prompt before generation.Source: Conclusion · p. 7
  • Reported results show competitive recommendation and weak-risk proxy-label prediction, while controlled Beauty analysis supports a contribution from graph conditioning.Source: Conclusion · p. 7
  • Routing analyses characterize group-dependent recommendation quality and inference cost but do not establish an advantage over All-Simple or matched-ratio random routing.Source: Conclusion · p. 7
Loading 2609.16560v1…