Source-linked AI summary
KB4Rec: A Dataset for Linking Knowledge Bases with Recommender Systems
Wayne Xin Zhao, Gaole He, Hongjian Dou, Jin Huang, Siqi Ouyang, Ji-Rong Wen
TL;DR
Existing RS side information is limited, and prior RS-to-KB linkage results may depend on private knowledge bases. KB4Rec v1.0 publicly links three RS datasets with Freebase and examines linkage patterns; linkage rates vary by dataset, and the analyses associate higher popularity and older item recency with greater likelihood of linkage.
Problem
Researchers seeking rich, structured knowledge for RS items face limited platform side information or linked datasets based on private KBs.
Source: §1 · p. 1Method
KB4Rec v1.0 heuristically links MovieLens, LFM-1b, and Amazon book items to Freebase entities using titles and domain-specific side information.
Source: §1 · p. 1Results
Takeaways & Limitations
The public linked ID pairs provide a resource for research-purpose studies and let researchers reuse Freebase information for RS items.
Source: §1 · p. 1, §3 · p. 2The recommendation comparison illustrates dataset use with only four methods; evaluating more knowledge-aware algorithms is left for future work.
Source: §5 · p. 3
Abstract
from arXiv · showhide
To develop a knowledge-aware recommender system, a key data problem is how we can obtain rich and structured knowledge information for recommender system (RS) items. Existing datasets or methods either use side information from original recommender systems (containing very few kinds of useful information) or utilize private knowledge base (KB). In this paper, we present the first public linked KB dataset for recommender systems, named KB4Rec v1.0, which has linked three widely used RS datasets with the popular KB Freebase. Based on our linked dataset, we first preform some interesting qualitative analysis experiments, in which we discuss the effect of two important factors (i.e. popularity and recency) on whether a RS item can be linked to a KB entity. Finally, we present the comparison of several knowledge-aware recommendation algorithms on our linked dataset.
1 INTRODUCTION
Knowledge-aware recommendation needs rich, structured item knowledge, but existing approaches either provide limited platform side information or rely on private KB linkages. KB4Rec v1.0 addresses this gap with a public dataset linking three RS datasets to Freebase, then examines linkage factors and recommendation algorithms.
1 INTRODUCTION
- Existing approaches offer limited attributes and relations from RS side information, while private-KB linkage results are not publicly available.These constraints leave researchers without a public linked dataset containing rich, structured knowledge for RS items.Source: §1 · p. 1
- KB4Rec v1.0 publicly links MovieLens, LFM-1b, and Amazon book items to entities in Freebase across movie, music, and book domains.The authors chose widely used RS datasets and Freebase, which is also used in NLP and related domains.Source: §1 · p. 1
- The dataset stores RS-item and KB-entity ID pairs, enabling researchers to retrieve linked entities’ Freebase relations and attributes.The paper also uses KB4Rec for qualitative analyses of popularity and recency and for comparing knowledge-aware recommendation algorithms.Source: §1 · p. 1
2 EXISTING DATASETS AND METHODS
Prior knowledge-aware recommendation work uses platform side information, HIN-based representations, or KBs, but some approaches have limited information or depend on private KBs. KB4Rec’s related-work distinction is its focus on linking RS items to public Freebase.
2 EXISTING DATASETS AND METHODS
- Platform side information in early context-aware recommendation includes social, POI, movie, and user-profile data, but often covers few information types and ignores relations among them.Examples include Epinions social networks, Yelp POI properties, MovieLens movie attributes, and microblogging user profiles.Source: §2 · pp. 1–2
- HINs organize side information through meta-path-based representations, but HIN-based recommendation can rely on graph search that is difficult to scale for relation-pattern discovery.Source: §2 · p. 2
- Earlier KB-based methods heuristically link RS items to KB entities but use private KBs, whereas this paper focuses on Freebase [18].The authors also note related linking studies.Source: §2 · p. 2
3 LINKED DATASET CONSTRUCTION
KB4Rec v1.0 links three RS datasets to Freebase using item titles and domain-specific side information as linkage constraints. Its linkage rates vary substantially across domains, and the released resource consists of linked ID pairs rather than the original datasets.
3 LINKED DATASET CONSTRUCTION
- The source datasets cover movies, music, and books; MovieLens 20M is the largest of four released MovieLens datasets by rating count.The selected data include MovieLens ratings, LFM-1b listening events, and Amazon book ratings.Source: §3 · p. 2
- Linkage queries Freebase with item titles, then refines same-title matches using movie IMDB IDs, music artist names, or book writer names.The process handles string-matching variations, discards items not accurately linked or rejected, and filters LFM-1b items with fewer than ten listening events.Source: §3 · p. 2
- MovieLens 20M had a 95.2% linkage ratio, with items accurately linked to a KB entity, compared with 19.4% for LFM-1b and 4.7% for the Amazon book domain.The authors suggest MovieLens 20M’s fewer, already-refined items and speculate that domain bias in Freebase may help explain the differences; despite low ratios, the latter two datasets had large absolute numbers of linked items.Source: §3 · p. 2, Table 1 · p. 2
- KB4Rec v1.0 contains 25,982 linked ID pairs for MovieLens 20M, 1,254,923 for LFM-1b, and 109,671 for Amazon book; each pair links an RS item ID to a KB entity ID.Source: §3 · p. 2
4 LINKAGE ANALYSIS
Linkage ratios differ with item popularity and recency in the reported analyses. Popularity’s association covers all three datasets, while the recency analysis is limited to MovieLens 20M because it alone supplies item release dates.
4 LINKAGE ANALYSIS
- Figure 1 groups the first three subfigures as popularity analyses and the last as recency analysis, with letters indicating ordered bin numbers.Source: Figure 1 · p. 3
Effect of Popularity on Linkage
- Higher-popularity bins have higher linkage ratios across the three datasets, indicating popularity is likely to have a positive effect on linkage.Popularity is measured by the number of users who interacted with an item; items are sorted and divided into five equal-sized bins.Source: §4 · p. 3
Effect of Recency on Linkage
- Linkage ratios gradually decrease across older-to-newer item bins in MovieLens 20M, indicating recency is likely to have a negative effect on linkage.Only MovieLens 20M has release-date information; the last bin’s sharp drop may reflect the dataset version’s April 2015 cutoff.Source: §4 · p. 3
5 EXPERIMENT
The experiments compare recommendation methods on linked datasets using a last-item prediction task. KSR performs best consistently across the three datasets, while BPR performs worst.
Experimental Setup
- Evaluation uses last-item recommendation: predict each user's final interaction from earlier interactions, ranking the true item among 100 negatives.The reported metrics are MRR, HR, and NDCG; the experiments use a recent-year LFM-1b subset, MovieLens 20M data from 2005–2015, and filtered datasets.Source: §5 · p. 3
KB Information Representation
- KB information is embedded in low-dimensional vectors using TransE, trained from linked entities and a one-step graph expansion after removing relations with fewer than 5,000 triples.Source: §5 · p. 3
Methods to Compare
Results and Analysis
- KSR works best consistently on all three datasets; mCKE beats SVDFeature only on LFM-1b.The authors suggest fixed KB embeddings in their mCKE implementation, unlike the original CKE model's adaptive updates, as a possible reason.Source: Table 2 · p. 4, §5 · p. 4
- BPR performs worst on all three datasets, while the other models perform better and incorporate KB information.Source: §5 · p. 4
- SVDFeature uses learned KB embeddings as context features, and each additional feature increases the model's parameters.Source: §5 · pp. 3–4
6 CONCLUSION
The paper introduces KB4Rec v1.0, a public dataset linking recommender-system items with knowledge-base entities across three domains. Future work will extend the linked datasets and consider other knowledge bases.
6 CONCLUSION
- KB4Rec v1.0 provides linked ID pairs across three domains; future work will link more recommender datasets with Freebase and consider YAGO and DBpedia.Source: §6 · p. 4