Source-linked AI summary
Adapting Knowledge Graphs for Behavior Denoising in Sequential Recommendation
Zichun Jin, Zihan Zhou, Yinan Liu, Bin Wang, Xiaochun Yang
TL;DR
Sequential recommendation must distinguish informative interactions from noisy behavior, but existing methods lack explicit relational evidence and raw KG connectivity is biased by structural exposure. AdaptedKG calibrates KG evidence using matched contexts and reference items, then uses offline retention coefficients to adapt training without changing inference. Across a standard recommender and multiple behavior-denoising models, the reported results show improved performance on every reported metric.
Problem
Existing behavior-denoising methods lack explicit inter-item relational evidence, while raw KG connectivity can bias reliability estimates toward popular, highly connected, or broadly covered items.
Method
AdaptedKG compares contexts with structurally matched alternatives to select distinctive paths, calibrates interaction support against matched references, and uses offline coefficients to gate histories and target losses.
Results
AdaptedKG achieves higher values for every reported metric across all four backbones, including a standard sequential recommender and three behavior-denoising models.
Takeaways & Limitations
Calibrated graph evidence can complement behavior denoising while leaving the recommendation backbone and inference procedure unchanged.
Takeaways & Limitations
Matched-background or matched-reference calibration is unavailable when n < 2 or when valid matched sets cannot be constructed.
Abstract
from arXiv · showhide
Sequential recommendation predicts the next item from a user's interaction history, but not every interaction is equally informative. Real logs combine persistent preferences with temporary needs, exploration, and incidental behavior, so some interactions can distort history representations or provide unreliable supervision. Existing denoising methods judge such interactions mainly from co-occurrence, order, or model predictions, without explicit evidence from relations between items. Knowledge graphs (KGs) offer this evidence, but item popularity, graph degree, uneven coverage, and widely shared entities can inflate connectivity and bias reliability estimates. Here we present AdaptedKG, which derives calibrated KG evidence for each training example without adding graph representations to the recommendation model. It first compares the observed context with structurally matched alternatives to identify relational paths that are unusually prominent and uses them to build a local KG view. It then compares each interaction with structurally matched reference items to calibrate its support within that view. The resulting retention coefficients gate historical representations and reweight target losses. All sample-specific scores are computed offline using training interactions and a fixed KG, so the backbone remains unchanged and no KG access is required at inference. Experiments show gains with a standard sequential recommender and multiple behavior-denoising sequential recommenders.
1. Introduction
Sequential recommendation histories mix reliable preferences with temporary, exploratory, and incidental behavior, while existing denoising methods lack explicit inter-item relational evidence. AdaptedKG addresses this gap by calibrating KG-based support against structurally matched contexts and references without changing the recommender backbone.
- Motivation: Interaction histories contain unevenly reliable evidence because real logs mix long-term preferences, short-term needs, exploration, and incidental behavior.Weakly associated interactions can disrupt historical representations and provide unstable training supervision.
- Motivation: Existing denoising methods mainly use co-occurrence, order, or model predictions, without explicit evidence from relations between items.
- Motivation: Knowledge-graph connectivity can bias reliability estimates because popularity, degree, coverage, and shared entities inflate the number of available paths.Relational support must therefore be assessed against structurally comparable references rather than raw connectivity alone.
- Approach: AdaptedKG compares observed contexts with structurally similar alternatives to identify prominent relational patterns and construct an example-specific local KG view.The first structural-matching stage characterizes background path distributions before retaining and weighting distinctive patterns.
- Approach: A second structural-matching stage calibrates interaction support using reference items, producing retention coefficients that modulate historical representations and target losses.Adaptation and sample-specific scoring run offline using training interactions and a fixed KG, leaving inference and the backbone unchanged.
2. Related Work
Prior behavior-denoising methods reduce unreliable sequential-recommendation behavior through learned sequence edits, bidirectional predictions, or interaction-derived relations.
- STEAM learns item-wise keep, delete, and insert operations from synthetically corrupted sequences.
- BirDRec uses bidirectional recommender predictions to rectify unreliable histories and targets with theoretical error guarantees.
- SSDRec exploits interaction-derived inter-sequence relations to guide self-augmentation and hierarchical denoising.
3. Methodology
AdaptedKG calibrates knowledge-graph evidence by comparing contexts and candidate interactions with structurally matched alternatives. The resulting retention coefficients support historical representation gating and target-loss reweighting, with offline computation and defined fallback conditions.
- Retention coefficients gate historical representations before sequence encoding and scale the target loss.
- Matched-null contexts replace positions with items similar in training popularity, KG degree, and linkage status while preserving length and order.This controls structural differences when identifying unusually prominent relational paths.
- Retained paths are those whose observed coverage exceeds the matched-null median, with higher null-relative rank receiving greater weight.The weighted paths with positive α_p(C) form the local path set P*.
- Matched reference items calibrate each candidate's retained-path support, assigning retention below one half for below-median rank and one otherwise.References are matched by training popularity, KG degree, and linkage status, and are scored after the retained path set is fixed.
- Sample-specific coefficients are computed offline from training interactions and a fixed KG, detached from recommendation optimization.When the context is empty, the KG channel sets R(v; C) = 1 and skips undefined quantities.
- The method requires at least two items and valid matched background or reference sets, while an empty retained path set is a boundary condition.
4. Experimental Results
On Steam Games, AdaptedKG’s matched-reference calibration reduces the association between retention estimates and item-side structural exposure, while improving recommendation metrics across four backbones. Ablations show that local adaptation and both matching stages contribute to the full model’s performance.
- Experimental setup: The Steam Games evaluation uses 25,389 users, 4,089 items, 328,278 interactions, and 462,016 KG triples across six relations.
- Experimental setup: The evaluation uses identical data splits, candidate sets, and backbone hyperparameters for each backbone and its AdaptedKG-enhanced counterpart.Retention coefficients and interaction-derived matching statistics use training interactions only.
- KG-side calibration: Matched-reference sampling produces substantially weaker correlations between retention coefficients and item-side structural exposure.The calibration varies only the reference sampler while fixing target queries, local path views, and reference counts.
- Recommendation performance: AdaptedKG achieves higher values for every reported metric across SASRec, STEAM, BirDRec, and SSDRec.The SASRec gains show compatibility with a standard sequential recommender, while the other gains indicate compatibility with interaction-log denoising mechanisms.
- Ablation study: Removing local adaptation, matched-null calibration, or matched-reference calibration lowers every reported metric relative to the full model.Removing local adaptation causes the largest drops, while removing either matching stage causes smaller declines.
5. Conclusion
AdaptedKG uses structurally matched contexts and reference items to calibrate knowledge-graph evidence for interactions. Its offline retention coefficients adjust training signals while leaving the backbone and inference procedure unchanged, and the results show that calibrated graph evidence can complement behavior denoising.
- AdaptedKG compares observed contexts with structurally similar alternatives to identify and weight relational paths that stand out from the background.
- It uses similar reference items to calibrate interaction support within the selected path evidence.
- The resulting retention coefficients adjust historical representations and target losses during training.
- All graph processing is completed offline, leaving the recommendation backbone and inference procedure unchanged.
- The results show that calibrated graph evidence can complement behavior denoising in sequential recommendation.
Declaration of Use of Generative AI
The authors used OpenAI Codex on a limited basis for language polishing and experimental-code implementation and debugging, with all assistance reviewed and revised.
- The authors used OpenAI Codex for limited language polishing and assistance with implementing and debugging experimental code.
- The authors critically reviewed, tested, and revised all tool-assisted text and code.
- The authors retained responsibility for the research design, methodological decisions, analysis, interpretation, and content.