PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Intents behind Interactions with Knowledge Graph for Recommendation

Xiang Wang, Tinglin Huang, Dingxian Wang, Yancheng Yuan, Zhenguang Liu, Xiangnan He, Tat-Seng Chua

arXiv:2102.07057v1cs.IRcs.AI

TL;DR

Existing GNN-based KG recommenders do not finely model multiple user intents or preserve relation dependencies along long-range paths. KGIN addresses these gaps with intent-aware relation combinations and relational path-aware aggregation, achieving significant improvements over state-of-the-art baselines while providing interpretable intent and path explanations.

  • Problem

    Existing GNN-based methods model user-item relations coarsely, overlooking multiple user intents and dependencies among relations in long-range connectivity.

  • Method

    KGIN models each intent as an attentive combination of KG relations, encourages intent independence, and recursively aggregates information along relational paths.

  • Results

    KGIN achieves significant improvements over state-of-the-art methods and provides interpretable explanations by identifying influential intents and relational paths.

  • Takeaways & Limitations

    Fine-grained intent modeling and relational path-aware aggregation improve recommendation performance and interpretability within the evaluated KG-based setting.

  • Takeaways & Limitations

    Intent granularity is dataset-dependent: overly fine-grained intents can impair accuracy, and the authors leave its further exploration to future work.

Abstract

from arXiv · show

Knowledge graph (KG) plays an increasingly important role in recommender systems. A recent technical trend is to develop end-to-end models founded on graph neural networks (GNNs). However, existing GNN-based models are coarse-grained in relational modeling, failing to (1) identify user-item relation at a fine-grained level of intents, and (2) exploit relation dependencies to preserve the semantics of long-range connectivity. In this study, we explore intents behind a user-item interaction by using auxiliary item knowledge, and propose a new model, Knowledge Graph-based Intent Network (KGIN). Technically, we model each intent as an attentive combination of KG relations, encouraging the independence of different intents for better model capability and interpretability. Furthermore, we devise a new information aggregation scheme for GNN, which recursively integrates the relation sequences of long-range connectivity (i.e., relational paths). This scheme allows us to distill useful information about user intents and encode them into the representations of users and items. Experimental results on three benchmark datasets show that, KGIN achieves significant improvements over the state-of-the-art methods like KGAT, KGNN-LS, and CKAN. Further analyses show that KGIN offers interpretable explanations for predictions by identifying influential intents and relational paths. The implementations are available at https://github.com/huangtinglin/Knowledge_Graph_based_Intent_Network.

1 INTRODUCTION

Existing KG-based recommenders improve relational modeling but remain coarse-grained about user intents and relational paths. KGIN addresses both limitations to improve recommendation capacity and interpretability.

  • KG enriches user-item modeling with entity and relation information that can improve recommendation accuracy and explainability.KG relations can reveal item relatedness and help interpret preferences, such as attributing a movie choice to its director.
  • Earlier methods supplement item representations with KG triplets, while later approaches use multi-hop paths to characterize user-item relations.The passage notes that path-based methods face issues including labor-intensive feature engineering and poor transferability.
  • Users may have multiple intents, such as combining director and star aspects for some movies and star and partner aspects for another.Ignoring these finer-grained intents limits the modeling of user-item interactions.
  • Node-based aggregation collects neighboring information without distinguishing the relational paths through which it arrives.This loses dependencies among relations that contribute to long-range connectivity.
  • KGIN models each intent as an attentive combination of KG relations and encourages different intents to remain independent.It also introduces relational path-aware aggregation to preserve long-range relational semantics and encode user-intent information in user and item representations.
  • KGIN reveals user intents, models long-range relational paths, and demonstrates superiority through empirical studies on three benchmark datasets.The contribution summary emphasizes both model capacity and interpretability.

2 PROBLEM FORMULATION

The paper formulates knowledge-aware recommendation from implicit user-item feedback and a heterogeneous knowledge graph. The task is to predict how likely a user is to adopt an item.

  • Interaction Data: The problem uses implicit feedback, where views, clicks, or purchases indicate prior user-item interactions.Observed feedback is represented as user-item pairs in O+.
  • Knowledge Graph: A knowledge graph stores real-world facts as heterogeneous entities and relations, including canonical and inverse directions.Each fact is represented by a head, relation, and tail triplet.
  • Knowledge Graph: Mappings between items and knowledge-graph entities allow the KG to profile items and complement interaction data.The KG includes information such as item attributes, taxonomy, and external commonsense knowledge.
  • Task Description: Given observed interactions O+ and knowledge graph G, the task is to learn a function predicting how likely a user is to adopt an item.

3 METHODOLOGY

KGIN represents user-item interactions through shared intents linked to KG relations and aggregates information along relational paths. Independence modeling makes intents more distinct and interpretable.

  • KGIN Overview: KGIN uses multiple latent intents to profile user-item relationships and combines intent modeling with relational path-aware aggregation.The model aims to preserve holistic semantics in long-range connectivity while producing user and item representations.
  • User Intent Modeling: An intent represents a reason for choosing an item and may capture combinations of attributes such as star and partner or director and genre.Different intents abstract different behavioral patterns shared across users.
  • User Intent Modeling: KGIN decomposes each observed user-item pair into one triplet per shared intent and organizes these triplets as an intent graph.This differs from the homogeneous collaborative graph used in earlier work.
  • User Intent Modeling: Each intent is modeled through an attentive combination of KG relation embeddings rather than being tied to a single relation.Attention scores quantify relation importance, and the weights capture common patterns across users.
  • Independence Modeling of Intents: Independence modeling reduces dependence among intent representations so different intents provide distinct information about user preferences.The paper discusses mutual information, Pearson correlation, and distance correlation as possible regularizers.

3.2 Relational Path-aware Aggregation

KGIN replaces node-based aggregation with intent-aware and relational path-aware message passing, preserving relation dependencies across multi-hop KG connectivity.

  • Relational Path-aware Aggregation: KGIN recursively aggregates multi-hop neighbors to integrate structural information into user and item representations.The GNN-based scheme builds representations layer by layer from neighboring users, items, and KG entities.
  • Motivation: Node-based aggregators lose path origins and explicit relation dependencies, limiting their ability to preserve structural information.They combine neighborhood information without distinguishing channels and mainly use relations as decay factors.
  • Aggregation Layer over Intent Graph: Intent-graph aggregation uses user history and intent connections to encode fine-grained behavioral patterns into user representations.Personalized attention differentiates the importance of each intent for a given user.
  • Aggregation Layer over Knowledge Graph: Knowledge-graph aggregation models each relation within the message, allowing the same entity to convey different meanings in different relational contexts.The relational message combines the relation embedding with the connected entity embedding through an element-wise product.
  • Capturing Relational Paths: For an l-hop path, KGIN represents connectivity through the sequence of relations and recursively integrates those relational signals.This path-aware representation captures interactions among relations and preserves holistic multi-hop semantics.

3.3 Model Prediction

KGIN predicts a user’s likelihood of adopting an item from the learned user and item representations.

  • Model Prediction: After L aggregation layers, KGIN sums user and item representations from different layers to form final representations.The resulting representations encode intent-aware relationships and KG relation dependencies.
  • Model Prediction: An inner product between the final user and item representations produces the item-adoption prediction score.The score estimates how likely the user is to adopt the item.

3.4 Model Optimization

KGIN optimizes recommendation scores with pairwise ranking while regularizing intent independence and model parameters.

  • Objective: Pairwise BPR loss trains KGIN to score each user’s historical items above unobserved items.Training pairs observed positive interactions with unobserved counterparts.
  • Objective: The objective combines BPR loss with an independence loss and L2 regularization.The independence term encourages distinct intent representations, while hyperparameters control the regularization terms.
  • Model Parameters: KGIN uses ID embeddings for users, KG entities, relations, and intents, together with attention weights as model parameters.The aggregation scheme discards nonlinear activation functions and feature transformation matrices.
  • Complexity: Aggregation costs scale as O(L|C|d) over the intent graph and O(L|G|d) over the knowledge graph.Here L is the number of layers, |C| and |G| count intent-graph and KG triplets, and d is the embedding size.

4 EXPERIMENTS

Experiments evaluate KGIN on three benchmark datasets against knowledge-aware recommendation baselines, using all-ranking top-K metrics and ablations of its relational modeling. KGIN consistently improves recommendation accuracy, while analyses examine relational depth, intent granularity, independence, and explainability.

  • Experimental Setup: Experiments use Amazon-Book, Last-FM, and Alibaba-iFashion, covering book, music, and fashion outfit recommendation.The evaluation uses a 10core setting and all-ranking evaluation over each user’s full set of previously unseen items.
  • Experimental Setup: KGIN is compared with MF, CKE, KGAT, KGNN-LS, CKAN, and RGCN across knowledge-aware recommendation settings.The baselines cover KG-free, embedding-based, and GNN-based methods.
  • Overall Performance: 14.51%, 13.97%, and 5.91% are KGIN’s ndcg@20 improvements over the strongest baselines on Amazon-Book, Last-FM, and Alibaba-iFashion, respectively.KGIN consistently outperforms all baselines across the three datasets and all reported measures.
  • Relational Modeling: Removing KG relations or user intents reduces predictive accuracy, showing that relational modeling and finer-grained intent modeling both contribute to KGIN.Without relations, the model loses relational semantics; without intents, it models only coarser-grained user preference.
  • Relational Modeling: Increasing aggregation depth generally improves results, but KGIN-3 is worse than KGIN-2 on Alibaba-iFashion because first-order connectivity dominates that dataset.KGIN-3 is consistently better than KGIN-2 on Amazon-Book and Last-FM, where higher-order connectivity is complementary to second-order connectivity.
  • Explainability: KGIN induces shared intents from KG-relation combinations and provides instance-wise explanations by identifying influential intents for individual interactions.For an Amazon-Book interaction, the influential intent explains the choice through interest in a featured artist and a certain version.

5 RELATED WORK

Existing KG-based recommender models use embeddings, paths, policies, or GNN aggregation to model user-item relationships and long-range connectivity. KGIN addresses limitations in GNN-based relational modeling by representing hidden user intents and preserving relational dependencies in paths.

  • Existing Methods: KG-based recommender models are grouped into embedding-based, path-based, policy-based, and GNN-based methods.These groups differ in whether they use first-order KG connectivity, extracted paths, learned path-finding policies, or recursive neighborhood aggregation.
  • Embedding-based Methods: Embedding-based methods learn entity embeddings from first-order interactions or KG triplets and use them as item prior or context information.Examples include applying TransE or TransH embeddings before matrix factorization or another recommender model.
  • Path-based Methods: Path-based methods extract user-item paths through KG entities, but recommendation accuracy depends heavily on path quality.These paths are processed with recurrent or memory networks to enhance user representations.
  • Policy-based Methods: Policy-based methods use reinforcement-learning agents to learn path-finding policies, but sparse rewards, large action spaces, and policy-gradient optimization make training difficult.The passage characterizes these methods as efficient alternatives to brute-force search while noting their optimization challenges.
  • GNN-based Methods: GNN-based methods recursively aggregate one-hop information to encode multi-hop connectivity, yet commonly assume one user-item relation and lose relational dependencies in paths.KGIN instead models relationships at intent granularity and encodes relational paths for performance and interpretability.

6 CONCLUSION AND FUTURE WORK

The paper concludes that KGIN improves relational modeling in knowledge-aware recommendation by uncovering intent-level user-item relationships and aggregating relational paths. It identifies sparse historical-interaction supervision as a limitation and proposes self-supervised and causal directions for future work.

  • Conclusion: KGIN models user-item relationships through KG-coupled intents and relational-path-aware aggregation to refine user and item representations.The framework targets both explainable intent semantics and relational information from multi-hop paths.
  • Conclusion: The paper analyzes KGIN’s effectiveness and explainability for knowledge-aware recommendation.This analysis accompanies the proposed framework’s two-dimensional relational modeling approach.
  • Future Work: Historical interactions provide sparse supervision for current KG-based recommendation models, limiting the quality of learned representations.Future work will investigate self-supervised tasks as auxiliary supervision.
  • Future Work: Future work proposes causal effect inference, counterfactual reasoning, and deconfounding to investigate and amplify biases in knowledge-aware recommendation.These directions are presented alongside self-supervised learning for uncovering internal relationships among data instances.

A.1 Reproducibility

The reproducibility appendix reports KGIN’s dataset-specific hyperparameter settings and states that the implementation artifacts have been released.

  • Hyperparameter Settings: Table 6 lists KGIN’s learning rate, embedding size, aggregation layers, number of user intents, independence coefficient, and L2 regularization coefficient.The listed settings are provided for three datasets.
  • Reproducibility: The authors released code, datasets, model parameters, and training logs to facilitate reproducibility.The release is hosted in the project’s GitHub repository.
Loading 2102.07057v1…