Source-linked AI summary
Learning Heterogeneous Knowledge Base Embeddings for Explainable Recommendation
Qingyao Ai, Vahid Azizi, Xu Chen, Yongfeng Zhang
TL;DR
Collaborative filtering systems commonly use unstructured data and have difficulty modeling explicit relationships among user and item information. This paper integrates CF with a heterogeneous user-item knowledge graph and uses soft matching over graph paths to generate personalized explanations. Experiments on real-world e-commerce datasets show it outperforms many state-of-the-art baselines.
Problem
Collaborative filtering systems using unstructured data have difficulty modeling explicit relationships among user and item information, while integrating structured behaviors and item properties remains a challenge.
Source: §1 · p. 1Method
The framework combines collaborative filtering with a graph of heterogeneous user behaviors and item knowledge, then soft-matches graph paths to generate personalized explanations.
Source: §1 · p. 2Results
The model consistently outperforms many state-of-the-art baselines on real-world e-commerce datasets.
Source: §1 · p. 2Takeaways & Limitations
The learned heterogeneous knowledge-base embeddings support personalized recommendations and explanation paths, and the approach can incorporate multiple relation types.
Source: §8 · p. 14Single-hop training may not guarantee high-quality matching across multiple hops, although it simplifies training and supports generalizability.
Source: §5.2 · p. 8
Abstract
from arXiv · showhide
Providing model-generated explanations in recommender systems is important to user experience. State-of-the-art recommendation algorithms - especially collaborative filtering (CF)-based approaches with shallow or deep models - usually work with various unstructured information sources for recommendation, such as textual reviews, visual images, and various implicit or explicit feedbacks. Though structured knowledge bases were considered in content-based approaches, they have been largely ignored recently due to the research focus on CF approaches. However, structured knowledge exhibit unique advantages in personalized recommendation systems. When the explicit knowledge about users and items is considered for recommendation, the system could provide highly customized recommendations based on users' historical behaviors and the knowledge is helpful for providing informed explanations regarding the recommended items. A great challenge for using knowledge bases for recommendation is how to integrate large-scale structured data, while taking advantage of collaborative filtering for highly accurate performance. Recent achievements in knowledge-base embedding (KBE) sheds light on this problem, which makes it possible to learn user and item representations while preserving the structure of their relationship with external knowledge for explanation. In this work, we propose to explain knowledge-base embeddings for explainable recommendation. Specifically, we propose a knowledge-base representation learning framework to embed heterogeneous entities for recommendation, and based on the embedded knowledge base, a soft matching algorithm is proposed to generate personalized explanations for the recommended items. Experimental results on real-world e-commerce datasets verified the superior recommendation performance and the explainability power of our approach compared with state-of-the-art baselines.
1. Introduction
Collaborative filtering systems using unstructured data have difficulty modeling explicit relationships among user and item information. The paper combines CF with a heterogeneous user-item knowledge graph and soft path matching to recommend items and generate explanations.
1. Introduction
- Knowledge-base embeddings can place heterogeneous information in a shared low-dimensional space while preserving its internal structure.The paper motivates this as a way to combine user behaviors and item properties in user and item representations.Source: §1 · p. 2
- The framework integrates multiple user behaviors and item knowledge in a unified graph, then extends collaborative filtering to learn over that heterogeneous knowledge.The authors say this captures user preferences more comprehensively.Source: §1 · p. 2
- Soft matching searches graph paths to construct personalized explanations for recommended items.The framework uses fuzzy reasoning over knowledge-graph paths for each recommended item.Source: §1 · p. 2
- Experiments on real-world e-commerce datasets show the model consistently outperforms many state-of-the-art baselines.Source: §1 · p. 2
2. Related Work
Prior work explored knowledge-based and semantic-graph recommendation, while later knowledge-base embedding methods offered a softer way to model relations. Earlier KBE recommenders did not preserve knowledge structure for reasoning or use it to generate explainable recommendations.
2. Related Work
- Earlier work discussed hybridizing knowledge-based and collaborative-filtering systems, building product-attribute recommenders, and using semantic interest graphs for cross-domain recommendations.Source: §2 · p. 2
- Later systems used linked open data, knowledge graphs for sound and music, and personalized page rank to recommend items with explanations.Source: §2 · pp. 2–3
- Reasoning over hard-coded heterogeneous graph paths limits prior approaches' ability to apply collaborative filtering across different entities and relations.Source: §2 · p. 3
- KBE represents entities and relations as vectors, enabling soft connectivity modeling; earlier KBE recommenders either collapsed item knowledge into one vector or omitted item and user knowledge.The paper also states that prior work had not used KBE for explainable recommendation [1].Source: §2 · p. 3
3. Problem Formulation
The paper formulates explainable product recommendation as retrieving likely purchases and explaining each user-item recommendation using a knowledge base of entity-relation triplets.
3. Problem Formulation
- The task is to retrieve products a user is likely to purchase and explain each user-item pair with a natural-language sentence based on the knowledge base.Source: §3 · pp. 3–4
- The formulation includes users, items, words, brands, and categories, connected by purchase, mention, category, producer, and item-to-item relations.The item-to-item relations include Bought_together, Also_bought, and Also_viewed.Source: §3 · pp. 3–4
4. Collaborative Filtering on Knowledge Graphs
The model extends collaborative filtering over a user-item knowledge graph by embedding entities and modeling relations as translations. It learns from observed triplets and uses relation-specific noise distributions during optimization.
4.1. Relation Modeling as Entity Translations
- Entities are embedded as vectors, while each relation translates a head entity toward a tail entity in the latent space.The framework models product knowledge as relational triplets.Source: §4.1 · p. 4
- Exact translations cannot represent entities linked to multiple distinct tails, so the model relaxes this constraint and assigns high probability to observed triplets and low probability to unobserved ones.The paper gives Also_viewed as an example of a one-to-many relation.Source: §4.1 · pp. 4–5
- A softmax over candidate tail entities turns translation into a probabilistic objective that supports soft matching for explanation extraction.The candidate set contains entities of the same type as the tail entity.Source: §4.1 · p. 5
4.2. Optimization Algorithm
- Because the full softmax is costly, training approximates observed-triplet likelihood with negative sampling and sums the resulting log-likelihood across triplets.Negative entities are sampled with the same type as the observed tail.Source: §4.2 · p. 5
- For relations other than Purchase, the noise distribution follows tail frequency; for Purchase, it is uniform to avoid bias toward particular items.The paper also interprets the resulting objective as jointly factorizing multiple relation-specific mutual-information matrices.Source: §4.2 · p. 6
5. Recommendation Explanation with Knowledge Reasoning
The section constructs recommendation explanations by finding paths between users and items in the knowledge graph, then ranking those paths using soft matching to select explanations for natural-language templates.
5. Recommendation Explanation with Knowledge Reasoning
- Breadth-first search from the user and item finds candidate paths up to depth z; the model scores and ranks them, returning the highest-probability path for an explanation.The paths can then be rendered with templates based on their relations.Source: §5 · pp. 6–7, §5.1 · p. 7
5.1. Explanation Path
- A path links a user and item when their relation sets can infer a shared entity; templates verbalize routes such as Bob and iPad sharing IOS or Apple.The Apple route uses Purchase + Produced_by on Bob’s side and Produced_by on the item’s side.Source: §5.1 · p. 7, Figure 2 · p. 7
5.2. Entity Soft Matching
- Because observed knowledge-graph triplets are sparse, latent-space entity soft matching constructs paths for arbitrary user–item pairs through an intermediate entity.The model computes path probabilities and ranks paths to select one for a natural-language explanation.Source: §5.2 · pp. 7–8
- Single-hop training simplifies the algorithm and supports generalizability, but may not ensure high-quality matching across multiple hops; direct multi-hop training is left for future work.The authors also state that the single-hop strategy competes with many baselines.Source: §5.2 · p. 8
6. Experimental Setup
The experiments use four Amazon 5-core product datasets and compare recommendation models with top-10 ranking metrics. The section also reports model-training and baseline-tuning settings.
6.1. Datasets
- The experiments use 5-core Amazon reviews for CDs & Vinyl, Clothing, Cell Phones, and Beauty, splitting purchases into 70% training and 30% testing.Each user or item has at least five associated reviews; users have at least three observed training reviews and two held-out reviews.Source: §6.1 · p. 8, Table 1 · p. 9
6.2. Evaluation
- The comparison includes matrix-factorization baselines BPR, BPR-HFT, and VBPR, and deep models DeepCoNN, CKE, and JRL; TransRec is also described as a baseline.The passage describes TransRec as using items as entities and users as relations, unlike this paper’s joint embedding of users, items, and other knowledge entities.Source: §6.2 · p. 9
- Evaluation uses NDCG, Precision, Recall, and Hit-Ratio on each test user’s top-10 recommendations, with significance tests based on Fisher randomization.Only purchased test-set items count as relevant.Source: §6.2 · p. 10
6.3. Parameter Settings
- The model trains for 20 epochs with batch size 64, five negative samples, and embedding dimensions tuned from 10 to 500; baseline hyperparameters are tuned by five-fold cross-validation.For BPR-HFT, 10 topics performs best; BPR and VBPR use λ = 10 in most cases.Source: §6.3 · p. 10
7. Results and Discussion
ECFKG outperforms baseline recommenders across four e-commerce datasets, and a case study illustrates explanation paths linking a user to a recommended item.
7.1. Recommendation Performance
- ECFKG improves over the best baseline, JRL, by 5.6% on CDs and Vinyl, 78.16% on Clothing, 23.05% on Cell Phones, and 45.56% on Beauty.The reported differences are significant at p < 0.001.Source: §7.1 · p. 10, Table 2 · p. 11
- ECFKG is robust to embedding-size variation; shallow models perform best at sizes 10–100, while neural models improve through size 300.Denser Cell Phones and Beauty datasets generally yield better ranking performance than the sparser CDs and Vinyl and Clothing datasets.Source: §7.1 · pp. 11–12, Figure 3 · p. 11
- ECFKG performs better than CKE, which the authors suggest may reflect CKE’s item-only user modeling and separate latent spaces that impede information propagation.The authors describe embedding-based relation modeling as a better way to incorporate structured knowledge for product recommendation.Source: §7.1 · pp. 11–12
- Modeling all relation types outperforms baselines and one- or two-relation variants, although the most useful added relation varies by product category.Mention helps most on CDs and Vinyl, whereas Also_view produces the largest improvement on Clothing.Source: §7.1 · p. 12, Table 3 · p. 12
- ECFKG takes about 10 hours to train on the largest dataset, comparable to deep models, while achieving better recommendation performance than other deep baselines.Generating a recommendation list for one user on that dataset takes less than 20 milliseconds.Source: §7.1 · pp. 12–13
7.2. Case Study for Explanation Generation
- For B9C, three explanation paths connect the user through co-purchased items or Chargers; the BTU path has the highest confidence at 7.39%.The user's training reviews mention iPad ownership, supporting the case study's interpretation of the co-purchased accessory recommendations.Source: §7.2 · pp. 13–14
8. Conclusions and Outlook
The paper concludes that heterogeneous knowledge-base embeddings can support personalized explainable recommendation. It argues that equipping recommender systems with knowledge remains important to improving performance and explainability.
8. Conclusions and Outlook
- The framework constructs a user-item knowledge graph from user behaviors and item knowledge, learns heterogeneous relations collectively, and uses user and item embeddings for personalized recommendations.Experimental results on real-world datasets show superior performance and flexibility to incorporate multiple relation types.Source: §8 · p. 14
- A soft matching algorithm finds explanation paths between users and recommended items in the latent knowledge-base embedding space.Source: §8 · p. 14
- The authors see equipping recommender systems with knowledge as important to improving performance and explainability in the future.Source: §8 · p. 14