PaperlayerView exact revision on arXiv

Source-linked AI summary

Unifying Knowledge Graph Learning and Recommendation: Towards a Better Understanding of User Preferences

Yixin Cao, Xiang Wang, Xiangnan He, Zikun hu, Tat-Seng Chua

arXiv:1902.06236v1cs.IR

TL;DR

Existing knowledge-graph recommender methods often assume incomplete graphs are complete, limiting the benefits of transferred knowledge and fine-grained preference understanding. This paper jointly models recommendation and KG completion with translation-based preference representations, improving both tasks while aligning preferences with KG relations for explainability.

  • Problem

    Existing KGs are often incomplete, so transferring their knowledge directly can limit recommendation benefits and obscure fine-grained user preferences.

  • Method

    The paper proposes TUP, a translation-based user-preference model, and jointly trains it with KG completion in KTUP by aligning latent preferences with KG relations.

  • Results

    Joint learning improves recommendation and KG completion, with KTUP achieving especially strong gains on sparse DBbook data and outperforming baseline methods across two datasets.

  • Takeaways & Limitations

    Aligning user preferences with knowledge-aware relations provides more explainable recommendations while coupling recommendation and KG completion.

  • Takeaways & Limitations

    The current model does not yet induce complex multi-hop preferences or apply KG reasoning to unseen preferences for cold-start users.

Abstract

from arXiv · show

Incorporating knowledge graph (KG) into recommender system is promising in improving the recommendation accuracy and explainability. However, existing methods largely assume that a KG is complete and simply transfer the "knowledge" in KG at the shallow level of entity raw data or embeddings. This may lead to suboptimal performance, since a practical KG can hardly be complete, and it is common that a KG has missing facts, relations, and entities. Thus, we argue that it is crucial to consider the incomplete nature of KG when incorporating it into recommender system. In this paper, we jointly learn the model of recommendation and knowledge graph completion. Distinct from previous KG-based recommendation methods, we transfer the relation information in KG, so as to understand the reasons that a user likes an item. As an example, if a user has watched several movies directed by (relation) the same person (entity), we can infer that the director relation plays a critical role when the user makes the decision, thus help to understand the user's preference at a finer granularity. Technically, we contribute a new translation-based recommendation model, which specially accounts for various preferences in translating a user to an item, and then jointly train it with a KG completion model by combining several transfer schemes. Extensive experiments on two benchmark datasets show that our method outperforms state-of-the-art KG-based recommendation methods. Further analysis verifies the positive effect of joint training on both tasks of recommendation and KG completion, and the advantage of our model in understanding user preference. We publish our project at https://github.com/TaoMiner/joint-kg-recommender.

1 INTRODUCTION

Knowledge graphs can improve recommendation accuracy and explainability, but their incompleteness can hide relevant relations. The paper jointly models recommendation and KG completion while transferring relation information to represent user preferences more explicitly.

  • Knowledge graphs encode entities and relations that can enrich recommender systems with structural information for accuracy and explainability.
  • Existing knowledge graphs are incomplete, so missing relations can prevent recommendation of relevant items despite correctly modeling a user’s observed preferences.The example omits the isDirectorOf relation between Robert Zemeckis and Death Becomes Her.
  • The paper jointly learns recommendation and KG completion so each task can support the other through shared information.KG facts augment user-item modeling, while enhanced interaction modeling helps complete missing facts.
  • KTUP transfers KG relation and entity embeddings into a translation-based user-preference model to give latent preferences explicit semantics.The model aligns preferences with relations such as director information and jointly learns users, items, entities, and relations.
  • Experiments on two datasets evaluate top-N recommendation and KG completion, with results reported as demonstrating the model’s effectiveness and explainability.

2 RELATED WORK

Related work spans similarity-based recommendation, neural models, side-information methods, KG-based recommendation, translation-based recommendation, and KG completion. The paper positions its approach around jointly using relational knowledge and user-item interactions.

  • Early recommendation methods used interaction histories to identify similar users or items, but extracting useful user and item features remained central.
  • Neural recommendation models automate latent feature extraction but continue to face data sparsity and cold-start problems.
  • Side information such as reviews, relational data, and knowledge graphs can address sparsity and improve explanations for recommendations.Explainability is associated with recommender effectiveness, efficiency, persuasiveness, and user satisfaction.
  • KG-based recommenders commonly augment user-item data with KG triplets or improve item embeddings using entity embeddings.Prior methods include factorization machines, unified graphs, and embedding-based approaches such as CoFM.
  • Translation-based recommendation models represent item selection through latent translational relations, while this paper addresses the N-to-N user-preference setting.N-to-N means one user may like multiple items and several users may like one item.
  • KG completion research learns low-dimensional entity and relation representations because existing knowledge graphs are incomplete.
  • TransE models a relation as a vector translation but struggles when one relation maps entities to multiple heads or tails.These cases include 1-to-N, N-to-1, and N-to-N relation patterns.
  • Similarity-based KG completion methods use semantic matching functions, including matrix, diagonal, and circular-correlation representations of relations.

3 PRELIMINARY

The preliminary section defines top-N recommendation and KG completion, introduces TUP’s preference-based scoring, and describes TransH as the KG-completion component. The framework addresses many-to-many relations through preference hyperplanes and relation-conditioned entity representations.

  • Tasks and Notations: Item recommendation uses implicit user-item feedback and aims to rank the top-N items for a target user.
  • Tasks and Notations: KG completion predicts a missing head or tail entity in a subject-property-object triple, equivalently ranking candidate entities for a query.
  • Tasks and Notations: TUP scores how likely a user is to like an item under an induced preference selected from a predefined preference set.Each preference acts similarly to a relation between two entities.
  • Tasks and Notations: KTUP can identify a director preference, recommend Death Becomes Her, and predict the missing entity in its isDirectorOf triplet.The example combines user-item interactions with structural KG knowledge.
  • KG Completion: TransE represents a valid triple by approximately satisfying e_h + r ≈ e_t, but one relation may correspond to multiple heads or tails.
  • KG Completion: TransH assigns each relation a hyperplane and validates translations after projecting the head and tail entities onto that shared hyperplane.
  • KG Completion: A lower TransH energy score indicates that a triple is more likely to be true.
  • KG Completion: The projected entity vectors and learned relation vectors define the translation score using an L1-norm distance.

4 TUP FOR ITEM RECOMMENDATION

TUP models user preferences as translational relationships between users and items, using preference induction and hyperplane-based translation to represent varied recommendation reasons.

  • TUP learns preference vectors p so that an interacted user-item pair approximately satisfies u + p ≈ i.
  • Preference Induction: Preference induction selects or combines latent factors that capture commonalities across users while complementing user-specific embeddings.The hard strategy selects one preference, whereas the soft strategy combines preferences through attention.
  • Preference Induction: The hard strategy uses ST Gumbel SoftMax to sample a discrete preference while retaining continuous gradients for end-to-end training.The forward pass uses a one-hot choice, while backward propagation follows the continuous distribution.
  • Preference Induction: The soft strategy addresses overlapping user motivations by combining multiple preferences according to attention weights based on similarity scores.
  • Hyperplane-based Translation: Hyperplane-based translation separates item representations and makes them similar only after projection onto a preference-specific hyperplane, addressing the N-to-N issue.This avoids treating every pair of items consumed for the same preference as universally interchangeable.
  • Hyperplane-based Translation: TUP trains recommendation distances with BPR loss so interacted items are closer than randomly corrupted non-interacted items.

5 JOINT LEARNING VIA KTUP FOR TWO TASKS

KTUP jointly learns recommendation and KG completion by transferring entity and relation knowledge into TUP, while using user preferences to improve KG modeling and interpretation.

  • KTUP extends TUP with KG entities and relations, jointly learning item recommendation and knowledge graph completion.The framework combines TUP for recommendation with TransH for KG completion.
  • Aligned entity embeddings enhance item embeddings, preserving structural KG knowledge that complements user-item interactions.For aligned item-entity pairs, the enhanced item embedding is i + e.
  • A predefined one-to-one mapping transfers relation embeddings into preference and projection vectors, linking KG relations to recommendation preferences.
  • Relation labels provide explicit explanations for preferences, such as isDirectorOf indicating a director preference.
  • KTUP combines user-item and KG learning through an overall objective with λ balancing the two tasks.

6 EXPERIMENTS

The experiments evaluate recommendation and KG completion across movie and book datasets, including sparsity, mutual task effects, and explainability analyses.

  • The study evaluates separate item-recommendation and KG-completion tasks on two domains while examining data sparsity, N-to-N behavior, joint training, and explainability.
  • Datasets: Items are mapped to DBPedia entities when mappings are available, and related facts are collected from directly connected DBPedia entities.The datasets use implicit positive feedback from ratings and randomly corrupted items as negatives.
  • Datasets: MovieLens-1m contains 6,040 users, 3,230 items, and 998,539 ratings after preprocessing.
  • Datasets: The authors note that unavailable released triplets required collecting their own KGs, producing slight differences from papers using the same datasets.
  • Datasets: DBbook2014 contains 5,576 users, 2,680 items, and 65,961 ratings, with a 99.6% sparsity rate.
  • Datasets: The constructed KGs contain 434,189 MovieLens triplets with 14,708 entities and 20 relations, and 334,511 DBbook triplets with 13,882 entities and 13 relations.

6.2 Baselines

The study compares recommendation and KG-completion methods under a shared experimental protocol, using standard baselines, fixed data splits, searched hyperparameters, and single training runs.

  • Baselines: Recommendation baselines include FM, BPRMF, CFKG, CKE, and CoFM, spanning similarity-based and KG-based approaches.
  • Baselines: KG completion comparisons include TransE, TransH, and TransR, along with evaluations of the KG-based recommendation methods.
  • Protocol: The authors reimplement baselines and remove unavailable reviews and visual-information components to ensure comparison on the selected datasets.
  • Protocol: Datasets are randomly split 7:1:2 into training, validation, and test sets, with at least one test item retained per user.
  • Protocol: Hyperparameters are selected through grid search for learning rate, L2 regularization, and optimization method, then applied across models.
  • Protocol: Experiments use batch size 256, embedding size 100, and early stopping on validation sets.
  • Protocol: The number of preferences and joint-loss weights are set using dataset relations and searches over λ values.
  • Protocol: Each model is trained once per dataset because the stated goal is measuring mutual task impacts rather than maximizing tuned performance.

6.4 Item Recommendation

The proposed methods achieve the strongest item-recommendation performance on both datasets. KG benefits are largest for moderately sparse users, while TUP is stronger for simpler preferences and KG effects diminish with abundant training data.

  • Metrics: Recommendation is evaluated with Precision@N, Recall@N, F1 score@N, Hit ratio@N, and nDCG@N over ranked candidate items.
  • The proposed methods perform best against baseline methods on both datasets.TUP is competitive without additional information, while KTUP gains 11.06% versus 4.43% in F1 on DBbook2014 and MovieLens-1m, respectively.
  • Influence of Training Data Sparsity: KG-based methods achieve their largest advantage when users average 100 to 200 training ratings.
  • Influence of Training Data Sparsity: The performance gap between KG-based and non-KG models narrows as training data decreases and nearly disappears at 563 average ratings.
  • Influence of Training Data Sparsity: TUP outperforms KTUP when users have fewer than 50 ratings, indicating that simpler preferences favor direct use of user-item interactions.

6.5 Knowledge Graph Completion

KTUP generally provides the strongest knowledge-graph completion results across datasets and relation categories. Its advantage is especially evident for complex N-to-N relations, while several transfer baselines can degrade KG completion.

  • Metrics: KG completion predicts a missing head or tail entity by ranking all entities as candidates, using Hit ratio and mean rank for evaluation.
  • Overall Results: KTUP almost outperforms all competing models on both datasets, except for mean rank on MovieLens-1m.
  • Overall Results: KTUP improves Hit Ratio over TransH by 2.67% on MovieLens-1m and 1.15% on DBbook2014.
  • Overall Results: CFKG, CKE, and CoFM underperform their basic KG components, potentially because recommendation constraints affect aligned entities and N-to-N preference representations.
  • Relation Categories: KTUP performs best on N-to-N relations and remains competitive with TransE and CoFM on 1-to-1, 1-to-N, and N-to-1 relations.

6.6 Mutual Benefits of Two Tasks

The training curves indicate mutual interaction between recommendation and KG completion, with KTUP and CFKG showing the strongest correlations. KTUP combines relation and preference structure while addressing complex relations through hyperplanes.

  • The correlation analysis uses training epoch on the x-axis, with KG-completion hit ratio and recommendation F1 on separate y-axes.
  • KTUP and CFKG show the strongest correlations between recommendation F1 and KG-completion hit-ratio curves.The two curves tend to increase and decrease simultaneously during training.
  • The results associate relation transfer with coordinated training of the two tasks.
  • KTUP combines relation and preference embeddings and introduces hyperplanes to address N-to-N issues.
  • CoFM and CKE show weaker curve correlations because only a small portion of entities is transferred.The text further characterizes CoFM as potentially unstable and CKE as unidirectionally enhancing recommendation at the expense of KG completion.

6.7 Case Study

The case study illustrates KTUP’s explainability by linking recommended movies to the user’s inferred preferences for director and starring relations.

  • KTUP identifies isDirectorOf and starring as the user’s strongest preferences among seven interacted movies.These preferences receive the highest attention in Section 4.1.
  • Batman Forever and Batman & Robin are recommended through the user’s preference for their director, Joel Schumacher.
  • The director preference also helps recommend Say Anything ..., directed by Cameron Crowe.
  • Starring preferences for James Dean and Natalie Wood contribute to recommending Rebel Without a Cause.

7 CONCLUSION

The paper concludes that KTUP extends TUP with KG completion, aligning user preferences with relations to improve explainability and both learning tasks. Future work targets more complex multi-hop preferences and KG reasoning for unseen preferences in cold-start settings.

  • TUP models multiple implicit user–item relations as preferences, while KTUP integrates this recommender with KG completion.
  • Future work will model multi-hop entity relations and use KG reasoning for unseen preferences to address cold-start problems.
Loading 1902.06236v1…