PaperlayerView exact revision on arXiv

Source-linked AI summary

Continual Graph Memory for Adaptive Recommendation under Intent Drift

Hao Nguyen Ngoc, Tung Nguyen, Nguyen Thi Hanh, Hoang Thai Dinh, Nguyen Xuan Tung

arXiv:2609.04651v1cs.AI

TL;DR

Adaptive recommendation under intent drift lacks a reliable way to update relational evidence while preserving stable knowledge. CGM-Rec makes the graph writable through quality-gated semantic edits and fast episodic lessons, achieving the strongest mean results on Bundle and consistently outperforming K-RagRec on metadata-rich ML-100K. Its evaluation remains bounded to sampled-candidate reranking, and episodic-to-semantic promotion may propagate biased or noisy evidence.

  • Problem

    Traditional KG-enhanced recommenders treat graphs as fixed retrieval sources, limiting adaptation when intents drift and feedback reveals noisy or missing relations.

  • Method

    CGM-Rec uses Episodic Lesson Memory for recent outcome-derived lessons and Semantic Graph Memory for conservative, quality-gated relational consolidation under frozen parameters.

  • Results

    CGM-Rec achieves the strongest mean results on Bundle and remains competitive on Games and ML-1M while consistently outperforming K-RagRec on metadata-rich ML-100K.

  • Takeaways & Limitations

    Writable graph memory provides a continual adaptation mechanism that separates rapid feedback reuse from stable semantic consolidation.

  • Takeaways & Limitations

    Evaluation focuses on sampled-candidate reranking rather than full-catalog evaluation, and episodic-to-semantic promotion may propagate biased or noisy evidence.

Abstract

from arXiv · show

This paper studies adaptive recommendation under intent drift, where feedback from each recommendation outcome can reveal whether the relational evidence used for ranking is useful, missing, or misleading. While Knowledge Graphs (KGs) provide essential semantic structure to handle these shifts, traditional KG-enhanced systems treat the graph as a static retrieval substrate, making it brittle to evolving intents, noisy metadata, and recurring failure patterns. This paper proposes CGM-Rec, a continual graph memory framework for adaptive recommendation. CGM-Rec treats the graph state as a writable memory and maintains two complementary components. Therein, a Semantic Graph Memory is updated conservatively through quality-gated typed operations for storing stable and high-confidence relational knowledge. Meanwhile, an Episodic Lesson Memory acts as a fast reactive memory that learns recent outcomes, failure cases, and corrective hints. During testing, model parameters remain frozen and adaptation occurs only through memory writes. We evaluate CGM-Rec under a frozen-parameter, one-pass reranking protocol, where encoders and prompts remain fixed during testing and adaptation occurs only through memory writes. Experiments across multiple recommendation settings show that CGM-Rec improves over evaluated neural and LLM-based baselines on most metrics. Particularly, under sampled-candidate reranking, CGM-Rec improves HR@1 by up to 29.58% over the strongest LLM baseline on Bundle, and outperforms K-RagRec on metadata-rich ML-100K with HR@5 of 0.5941 versus 0.4746.

1 Introduction

Under intent drift, conventional recommenders and read-only knowledge graphs struggle to update which relational evidence remains useful. CGM-Rec addresses this by treating graph state as writable memory with fast episodic feedback and conservative semantic consolidation.

  • Intent drift requires recommenders to decide which evidence remains trusted as outcomes and item semantics change.
  • Read-only KG-enhanced systems keep retrieving noisy relations and cannot add or reinforce useful relations revealed by feedback.
  • CGM-Rec treats graph state as a writable memory updated after recommendation outcomes are observed.
  • Episodic Lesson Memory rapidly records recent feedback and correction hints, while Semantic Graph Memory stores structural knowledge for long-term robustness.
  • Under frozen parameters and a strict one-pass inference protocol, CGM-Rec outperforms state-of-the-art baselines on most evaluated metrics.

2 Related Work

Prior recommendation methods capture sequential patterns, exploit external graph structure, or adapt through prompts and memories, but generally do not maintain structured graph memories that can be corrected after outcomes. CGM-Rec instead constrains long-term adaptation to quality-gated graph-memory edits.

  • Session-based and sequential recommenders learn transition patterns, but their adaptive state is mainly encoded in parameters or hidden sequence representations.
  • These sequence-based methods do not explicitly maintain external relational knowledge that can be corrected, consolidated, or pruned after outcomes.
  • KG-enhanced recommenders enrich item representations and connect sparse interactions by propagating preferences over structured relations.
  • LLM recommenders adapt through prompts, textual memories, or agent states, while CGM-Rec uses a frozen recommender and analyzer.
  • CGM-Rec differs from dynamic memory methods by constraining persistent adaptation to quality-gated graph-memory edits.

3 Task Definition

The task is one-pass candidate reranking over an ordered stream, where a dual-timescale memory informs each ranking and feedback updates memory afterward. Parameters and thresholds remain frozen, while a finite budget maintains bounded memory through deterministic eviction.

  • Adaptive recommendation is formulated as one-pass reranking over ordered instances containing context, candidate pool, and ground-truth item.
  • The recommender ranks each candidate pool using context and dual-timescale Semantic Graph and Episodic Lesson memories.
  • After evaluating the ranking against the target, a deterministic write policy updates memory, while recommender parameters and thresholds remain frozen.
  • The objective is to maximize stream-level ranking quality using metrics such as HR@K and NDCG@K.
  • A strict finite memory budget triggers deterministic eviction of stale, low-utility, or conflicting records when capacity is reached.

4 Methodology

CGM-Rec makes graph state the adaptive object by combining a fast episodic lesson memory with a conservatively updated semantic graph memory. Under frozen parameters, it retrieves both evidence types, analyzes revealed outcomes, and writes only quality-approved memory updates.

  • 4.2 Dual-Memory Architecture: CGM-Rec uses Semantic Graph Memory for stable relational knowledge and Episodic Lesson Memory for recent outcome-derived lessons and correction signals.ELM adapts quickly, while SGM changes conservatively through gated graph edits.
  • 4.2 Dual-Memory Architecture: The seed graph initializes SGM and supports a relation-aware graph encoder that is trained offline and frozen before inference.At inference, the encoder retrieves graph evidence but does not update parameters.
  • 4.3 Dual-memory recommendation: Before reranking, the system retrieves local graph evidence from SGM and top-k contextually relevant lessons from ELM, then fuses them in a structured prompt.SGM retrieval uses the frozen graph encoder, whereas ELM retrieval uses BM25 over a query built from context, candidates, and attributes.
  • 4.4 Quality-Gated Memory Update: After ranking, the revealed target determines a rank-based success or failure label that drives lesson generation and memory updates.The target is withheld until after ranking, preserving the predict–reveal–update sequence.
  • 4.4 Quality-Gated Memory Update: The Analyzer produces lessons and structured edit proposals, but only proposals passing a deterministic quality gate can modify SGM.The gate checks rank, support, confidence, conflicts, memory cost, and recency before accepted typed actions are applied.
  • 4.4 Quality-Gated Memory Update: Accepted graph edits reinforce useful relations, suppress misleading ones, insert tentative low-confidence edges, and prune weak or unsupported edges.These operations aim to preserve useful structure while limiting noise accumulation over long streams.

5 Experiments

The experiments evaluate CGM-Rec on four constrained candidate-reranking benchmarks against traditional, neural, and LLM-based baselines under continual predict–reveal–update evaluation. The setup fixes model backbones and controls leakage while measuring ranking quality and efficiency.

  • 5 Experiments: The study addresses whether CGM-Rec outperforms traditional, neural, and LLM-based baselines, especially in metadata-rich contexts and under continual intent shifts.It also examines the contributions of dual memory and graph updating.
  • 5.1 Experimental Setup: CGM-Rec is evaluated on Bundle, Games, ML-1M, and metadata-rich ML-100K using candidate sets containing the target and sampled items.The evaluation measures constrained candidate reranking rather than full-catalog retrieval.
  • 5.1 Experimental Setup: All continual evaluations follow a strict predict–reveal–update protocol, with target feedback exposed only after each ranking is produced.Main comparisons are repeated over five matched seeds and reported as mean ± standard deviation.
  • 5.1 Experimental Setup: The evaluation uses HR@K and NDCG@K for K ∈{1, 5, 10}, where HR measures top-K inclusion and NDCG additionally accounts for ranking position.Higher values indicate better ranking performance.
  • 5.1 Experimental Setup: The benchmark includes popularity-based, session-based, neural, graph-enhanced, and LLM-based recommenders, with all LLM methods using GPT-4.1-mini on candidate sets of size 20.This design keeps backbone and candidate-set conditions aligned across LLM comparisons.

6 Experiment Results

Across sampled-candidate reranking and continual Bundle evaluation, CGM-Rec generally outperforms the evaluated baselines, while regime-shift recovery remains incomplete.

  • Overall Performance Comparison: CGM-Rec achieves the best results on most datasets and metrics, with significant Bundle gains over the runner-up but nonsignificant differences on Games and ML-1M.Table 2 reports mean ± SD over five seeds; Bundle differences against the runner-up are significant for HR@1 and HR@5.
  • Overall Performance Comparison: 0.3263 HR@1 is achieved by CGM-Rec on Bundle, compared with 0.0630 for Atten-Mixer.The comparison is reported against the strongest neural baseline on Bundle.
  • LLM-based Comparison in Metadata-rich: +0.0971 HR@1 is the gain over K-RagRec on metadata-rich ML-100K, with all gains significant at p < 0.05.The comparison targets the strongest LLM graph-augmented baseline in the non-session setting.
  • LLM-based Comparison in Metadata-rich: CGM-Rec maintains higher mean HR@5 than K-RagRec across candidate pools C ∈{20, 50, 80} on both Games and ML-100K.These are candidate-pool scaling robustness tests.
  • Continual Adaptation: CGM-Rec outperforms all baselines throughout the Bundle stream, and its later advantage exceeds the ±1 SD bands.The curve increases in later snapshots while memory updates occur every five sessions and parameters remain frozen.
  • Adaptation under Intent-Regime Shifts: CGM-Rec recovers a larger proportion of lost HR@5 after regime transitions than Static SGM+ELM and K-RagRec, but recovery remains partial.Later HR@5 remains 0.050 and 0.040 below pre-shift levels after the two transitions.

7 Conclusion

The conclusion presents CGM-Rec as a writable graph-memory framework that separates rapid feedback capture from stable semantic consolidation under frozen-parameter reranking.

  • Conclusion: CGM-Rec moves KG-enhanced recommendation beyond read-only retrieval by treating graph state as writable memory.Its quality-gated writer converts recommendation outcomes into typed, provenance-aware edits.
  • Conclusion: Episodic Lesson Memory captures fast feedback, while Semantic Graph Memory performs stable semantic consolidation.The conclusion attributes complementary benefits to episodic lessons and semantic graph memory.
  • Conclusion: Under frozen-parameter, one-pass sampled-candidate reranking, CGM-Rec achieves the strongest mean results on Bundle and remains competitive on Games and ML-1M.It consistently outperforms K-RagRec on metadata-rich ML-100K.

8 Limitations

The paper’s limitations concern operational cost, incomplete exploration of update policies, sampled-candidate scope, the controlled drift setting, and possible noise propagation.

  • Limitations: CGM-Rec requires extra storage and computation for memory maintenance compared with static retrieval.This adds operational complexity to the adaptive architecture.
  • Limitations: Broader quality-gate configurations and typed update magnitudes remain unexplored.Robustness was tested only under moderate threshold perturbations.
  • Limitations: Evaluation focuses on sampled-candidate reranking rather than full-catalog evaluation.C = 50 and C = 80 are treated as larger-pool robustness tests.
  • Limitations: The controlled Bundle analysis represents stream-level intent-regime shifts, not natural within-user preference drift.Bundle does not preserve persistent user identities across sessions.
  • Limitations: Episodic-to-semantic promotion may propagate biased or noisy evidence, motivating stronger provenance safeguards.This is a stated risk of converting episodic lessons into longer-term graph knowledge.

E Runtime and Token Overhead

CGM-Rec avoids test-time parameter updates but incurs overhead from two LLM calls per recommendation, while memory retrieval and updates add approximately 0.2 seconds per instance.

  • Runtime overhead: Two LLM calls dominate CGM-Rec’s per-instance runtime.Memory retrieval and update operations do not invoke the LLM.
  • Runtime overhead: Approximately 0.2 s per instance is added by memory retrieval and update.
  • Trade-off: CGM-Rec avoids test-time parameter updates while requiring two LLM calls per recommendation.
  • Evaluation protocol: The shared sampled-candidate protocol compares methods on an identical reranking space, whereas full-catalog numbers are only a neural-baseline sanity check.The full catalog contains 3,416 eligible items, and restricting its score vector to 20 candidates reproduces exact sampled-20 target ranks.

G Detailed Ablation Study (RQ3)

The ablation study separates the contribution of dual-memory retrieval from feedback-driven graph updates. Dual-memory use improves mean performance beyond either memory alone, while graph writes provide an additional dataset-dependent gain, especially on ML-100K.

  • Memory composition: Dual-memory retrieval yields higher mean performance than either Semantic Graph Memory or Episodic Lesson Memory alone.
  • Feedback-driven updates: HR@5 on ML-100K increases from 0.4460 to 0.5941 when feedback-driven graph updates are enabled.The write contribution is modest on Bundle and Games, more noticeable on ML-1M, and most pronounced on ML-100K.
  • Feedback-driven updates: Feedback-driven semantic-graph writes provide an additional source of adaptation whose magnitude varies by dataset.
  • Experimental setup: The evaluation uses established public benchmarks and processed recommendation episodes rather than newly collected user data.Dataset statistics distinguish source interactions from processed train and test episodes and define eligible items as the post-filter ranking vocabulary.

H Quality-Gate Sensitivity Analysis

A one-factor-at-a-time analysis varies the quality-gate acceptance threshold and outcome-signal coefficient while holding the remaining system configuration fixed. Recommendation performance remains relatively stable despite substantial changes in accepted-update rates.

  • Sensitivity results: Varying either quality-gate parameter substantially changes the accepted-update rate while recommendation performance remains relatively stable.The analysis varies ηsup and θout independently with prompts, update rules, candidate sets, memory budgets, and initial states fixed.
  • Sensitivity results: The sensitivity analysis captures end-to-end effects because accepted updates alter future memory states and retrieval.
  • Robustness: The default configuration is retained without post-hoc selection, supporting robustness to moderate gate perturbations.

I Candidate-Pool Scaling Analysis

CGM-Rec retains higher mean performance than K-RagRec as sampled candidate pools grow from 20 to 80 items, although absolute ranking performance declines with harder negatives. The evidence supports robustness within sampled reranking, not full-catalog recommendation or training-scale scalability.

  • Scaling results: CGM-Rec maintains higher mean performance than K-RagRec across all evaluated metrics and candidate-pool sizes.The comparison uses identical contexts, candidate files, instance order, and five matched candidate-set seeds.
  • Scaling results: HR@5 relative gains on Games increase from 6.3% at C = 20 to 18.1% at C = 80.
  • Scaling results: ML-100K retains HR@5 gains of 15.2–25.2% across the tested candidate pools.The HR@5 improvements are statistically significant at every candidate size on both datasets.
  • Scaling results: Absolute ranking performance declines as candidate pools become larger and more challenging.The pools are nested, preserving original candidates while adding harder negatives.
  • Scope: The analysis measures constrained sampled reranking rather than full-catalog recommendation or training-scale scalability.
  • Mechanism: CGM-Rec’s gains are attributed primarily to explicit memory maintenance, with ELM reusing recent lessons and SGM preserving stable relational knowledge through conservative edits.
Loading 2609.04651v1…