An Empirical Comparison of Knowledge Graph Embeddings for Item Recommendation
4 likes1,127 views
The document discusses an empirical comparison of knowledge graph embeddings for item recommendation, highlighting the use of hybrid recommender systems that integrate content-based, collaborative filtering, and knowledge graph approaches. It presents various translational models, namely TransH, TransR, and TransE, and their performance metrics against baseline methods. Results indicate that TransH performs the best in terms of recommendation quality, emphasizing the potential of knowledge graphs in enhancing recommendation systems.
![FEATURES IN THE EMBEDDING SPACE
F
Feature engineering
● Manually define a mapping
function
Feature learning
● Define F (W) as a function of
a set of parameters W (e.g. a
neural network)
● Learn W from an
optimization problem
Random walks:
● DeepWalk [1]
● Node2vec [2]
● RDF2Vec [3]
● Entity2rec [4]](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-12-320.jpg)
![FEATURES IN THE EMBEDDING SPACE
F
Feature engineering
● Manually define a mapping
function
Feature learning
● Define F (W) as a function of
a set of parameters W (e.g. a
neural network)
● Learn W from an
optimization problem
Random walks:
● DeepWalk [1]
● Node2vec [2]
● RDF2Vec [3]
● Entity2rec [4]
Translational models:
● TransE [5]
● TransH [6]
● TransR [7]](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-13-320.jpg)
![FEATURES IN THE EMBEDDING SPACE
F
Feature engineering
● Manually define a mapping
function
Feature learning
● Define F (W) as a function of
a set of parameters W (e.g. a
neural network)
● Learn W from an
optimization problem
Random walks:
● DeepWalk [1]
● Node2vec [2]
● RDF2Vec [3]
● Entity2rec [4]
Translational models:
● TransE [5]
● TransH [6]
● TransR [7]](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-14-320.jpg)
![TRANSLATIONAL MODELS
TransE [5]
Model: h + r ~ t
Simple, but not suitable
for 1-to-N, N-to-1, N-to-N
relations](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-15-320.jpg)
![TRANSLATIONAL MODELS
TransH [6]
Model: h⊥
+ dr
~ t⊥
Multiple representations
for different relations
projecting on hyperplanes
TransE [5]
Model: h + r ~ t
Simple, but not suitable
for 1-to-N, N-to-1, N-to-N
relations](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-16-320.jpg)
![TRANSLATIONAL MODELS
TransR [7]
Model: hr
+ r ~ tr
Embed entities and relations
in different vector spaces
through projection matrix Mr
TransH [6]
Model: h⊥
+ dr
~ t⊥
Multiple representations
for different relations
projecting on hyperplanes
TransE [5]
Model: h + r ~ t
Simple, but not suitable
for 1-to-N, N-to-1, N-to-N
relations](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-17-320.jpg)
![TRANSLATIONAL MODELS
TransR [7]TransH [6]TransE [5]
Score function:
f(h, r, t) = D(h + r, t)
Score function:
f(h, r, t) = D(h⊥
+ dr
, t⊥
)
Score function:
f(h, r, t) = D(hr
+ r , tr
)](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-18-320.jpg)
![RANKING FUNCTION
i1
i3
ρ(u,i) = - D(u + feedback, i)
i2
i3
TransE [5]](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-19-320.jpg)
![RANKING FUNCTION
i1
i3
ρ(u,i) = - D(u + feedback, i)
i2
i3
TransE [5]
TransH [6]
ρ(u,i) = - D(u⊥
+ dfeedback
, i⊥
)](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-20-320.jpg)
![RANKING FUNCTION
i1
i3
ρ(u,i) = - D(u + feedback, i)
i2
i3
TransE [5]
TransH [6]
ρ(u,i) = - D(u⊥
+ dfeedback
, i⊥
)
TransR [7]
ρ(u,i) = - D(ufeedback
+ feedback, ifeedback
)
D = Euclidean distance](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-21-320.jpg)
![DATASET
● Movielens_1M: 1,000,209 ratings, 3900 movies, 6040 users
● Splitting: we split data using 70% train, 10% validation, 20% test of ratings per user
● DBpedia: we leverage DBpedia mappings of Movielens items to build the knowledge
graph created in a related work1
● Property selection: we count the most frequently occurring properties used in
DBpedia for items in Movielens1M, obtaining: ["dbo:director", "dbo:starring",
"dbo:distributor", "dbo:writer","dbo:musicComposer",
"dbo:producer","dbo:cinematography", "dbo:editing"]
● We add: "dct:subject" to this set and we model ratings >= 4 as “feedback” property
1: https://github.com/sisinflab/LODrecsys-datasets](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-22-320.jpg)
![EXPERIMENTAL SETUP
● Evaluation Protocol: we apply the same candidate generation strategies to all
systems under evaluation, i.e. “AllUnratedItems” [8]
● Baselines: we use heuristics such as MostPopular and collaborative filtering
algorithms such as Singular Value Decomposition (SVD) and ItemKNN from the
Surprise Library (http://surpriselib.com)
● Scoring: we measure standard information retrieval metrics such as P@5, P@10,
MAP, NDCG. The same scorer is applied to all systems
● Configuration: all systems are used with their default hyper-parameters
● Implementation of translational models is: https://github.com/thunlp/KB2E
● Evaluators, scorers, baselines and python wrappers of translational models can be
found at: https://github.com/D2KLab/entity2rec/tree/dev](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-23-320.jpg)
![REFERENCES
● [1]: Perozzi, Bryan, Rami Al-Rfou, and Steven Skiena. "Deepwalk: Online learning of social representations." Proceedings of the
20th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2014.
● [2]: Grover, Aditya, and Jure Leskovec. "node2vec: Scalable feature learning for networks." Proceedings of the 22nd ACM
SIGKDD international conference on Knowledge discovery and data mining. ACM, 2016.
● [3]: Ristoski, Petar, and Heiko Paulheim. "Rdf2vec: Rdf graph embeddings for data mining." International Semantic Web
Conference. Springer, Cham, 2016.
● [4]: Palumbo, Enrico, Giuseppe Rizzo, and Raphaël Troncy. "Entity2rec: Learning user-item relatedness from knowledge graphs
for top-n item recommendation." Proceedings of the Eleventh ACM Conference on Recommender Systems. ACM, 2017.
● [5]: Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., & Yakhnenko, O. (2013). Translating embeddings for modeling
multi-relational data. In Advances in neural information processing systems (pp. 2787-2795).
● [6]: Wang, Z., Zhang, J., Feng, J., & Chen, Z. (2014, July). Knowledge Graph Embedding by Translating on Hyperplanes. In AAAI
(Vol. 14, pp. 1112-1119).
● [7]: Lin, Yankai, et al. "Learning entity and relation embeddings for knowledge graph completion." AAAI. Vol. 15. 2015.
● [8]: Harald Steck. 2013. Evaluation of recommendations: rating-prediction and ranking. In Proceedings of the 7th ACM
conference on Recommender systems. ACM, 213–220.](https://image.slidesharecdn.com/comparisonknowledgegraphembeddingsitemrecommendation-180605065310/85/An-Empirical-Comparison-of-Knowledge-Graph-Embeddings-for-Item-Recommendation-27-320.jpg)
An Empirical Comparison of Knowledge Graph Embeddings for Item Recommendation
- 1. An Empirical Comparison of Knowledge Graph Embeddings for Item Recommendation Enrico Palumbo, Giuseppe Rizzo, Raphael Troncy, Elena Baralis, Michele Osella, Enrico Ferro
- 3. ITEM RECOMMENDATION USER ITEMS ranking function ρ (u, i) 0.8 0.6 0.5 0.2ρ(u,i)
- 4. CONTENT-BASED FILTERING JANE ρ (u, i) 0.8Kill_Bill_Vol.2 Samuel_Jackson 0.0 Content similarities with items liked in the past
- 5. COLLABORATIVE FILTERING ρ (u, i) 0.8 0.0 collaborative: users who watch x also like y... Kill_Bill_Vol.2 Jane Mark Taxi_Driver
- 6. BEYOND CONTENT-BASED AND COLLABORATIVE... Content-based ● overspecialization ● new users ● requires item metadata Collaborative ● new users ● new items ● data sparsity Hybrid recommender systems
- 7. Knowledge Graph ● K = (E, R, O) ● Entities E ● Users U ⊂ ● Items I ⊂ ● ● ● ⊂ ● feedback ⊂ Kill_Bill_Vol.2 Samuel_Jackson JaneMark Quentin_Tarantino Taxi_Driver Star_Wars_Ep.1 THE_AVENGERS
- 8. RECOMMENDATIONS USING A KNOWLEDGE GRAPH Item recommendation as a knowledge graph completion problem!
- 10. FEATURES IN THE EMBEDDING SPACE F Feature engineering ● Manually define a mapping function
- 11. FEATURES IN THE EMBEDDING SPACE F Feature engineering ● Manually define a mapping function Feature learning ● Define F (W) as a function of a set of parameters W (e.g. a neural network) ● Learn W from an optimization problem
- 12. FEATURES IN THE EMBEDDING SPACE F Feature engineering ● Manually define a mapping function Feature learning ● Define F (W) as a function of a set of parameters W (e.g. a neural network) ● Learn W from an optimization problem Random walks: ● DeepWalk [1] ● Node2vec [2] ● RDF2Vec [3] ● Entity2rec [4]
- 13. FEATURES IN THE EMBEDDING SPACE F Feature engineering ● Manually define a mapping function Feature learning ● Define F (W) as a function of a set of parameters W (e.g. a neural network) ● Learn W from an optimization problem Random walks: ● DeepWalk [1] ● Node2vec [2] ● RDF2Vec [3] ● Entity2rec [4] Translational models: ● TransE [5] ● TransH [6] ● TransR [7]
- 14. FEATURES IN THE EMBEDDING SPACE F Feature engineering ● Manually define a mapping function Feature learning ● Define F (W) as a function of a set of parameters W (e.g. a neural network) ● Learn W from an optimization problem Random walks: ● DeepWalk [1] ● Node2vec [2] ● RDF2Vec [3] ● Entity2rec [4] Translational models: ● TransE [5] ● TransH [6] ● TransR [7]
- 15. TRANSLATIONAL MODELS TransE [5] Model: h + r ~ t Simple, but not suitable for 1-to-N, N-to-1, N-to-N relations
- 16. TRANSLATIONAL MODELS TransH [6] Model: h⊥ + dr ~ t⊥ Multiple representations for different relations projecting on hyperplanes TransE [5] Model: h + r ~ t Simple, but not suitable for 1-to-N, N-to-1, N-to-N relations
- 17. TRANSLATIONAL MODELS TransR [7] Model: hr + r ~ tr Embed entities and relations in different vector spaces through projection matrix Mr TransH [6] Model: h⊥ + dr ~ t⊥ Multiple representations for different relations projecting on hyperplanes TransE [5] Model: h + r ~ t Simple, but not suitable for 1-to-N, N-to-1, N-to-N relations
- 18. TRANSLATIONAL MODELS TransR [7]TransH [6]TransE [5] Score function: f(h, r, t) = D(h + r, t) Score function: f(h, r, t) = D(h⊥ + dr , t⊥ ) Score function: f(h, r, t) = D(hr + r , tr )
- 19. RANKING FUNCTION i1 i3 ρ(u,i) = - D(u + feedback, i) i2 i3 TransE [5]
- 20. RANKING FUNCTION i1 i3 ρ(u,i) = - D(u + feedback, i) i2 i3 TransE [5] TransH [6] ρ(u,i) = - D(u⊥ + dfeedback , i⊥ )
- 21. RANKING FUNCTION i1 i3 ρ(u,i) = - D(u + feedback, i) i2 i3 TransE [5] TransH [6] ρ(u,i) = - D(u⊥ + dfeedback , i⊥ ) TransR [7] ρ(u,i) = - D(ufeedback + feedback, ifeedback ) D = Euclidean distance
- 22. DATASET ● Movielens_1M: 1,000,209 ratings, 3900 movies, 6040 users ● Splitting: we split data using 70% train, 10% validation, 20% test of ratings per user ● DBpedia: we leverage DBpedia mappings of Movielens items to build the knowledge graph created in a related work1 ● Property selection: we count the most frequently occurring properties used in DBpedia for items in Movielens1M, obtaining: ["dbo:director", "dbo:starring", "dbo:distributor", "dbo:writer","dbo:musicComposer", "dbo:producer","dbo:cinematography", "dbo:editing"] ● We add: "dct:subject" to this set and we model ratings >= 4 as “feedback” property 1: https://github.com/sisinflab/LODrecsys-datasets
- 23. EXPERIMENTAL SETUP ● Evaluation Protocol: we apply the same candidate generation strategies to all systems under evaluation, i.e. “AllUnratedItems” [8] ● Baselines: we use heuristics such as MostPopular and collaborative filtering algorithms such as Singular Value Decomposition (SVD) and ItemKNN from the Surprise Library (http://surpriselib.com) ● Scoring: we measure standard information retrieval metrics such as P@5, P@10, MAP, NDCG. The same scorer is applied to all systems ● Configuration: all systems are used with their default hyper-parameters ● Implementation of translational models is: https://github.com/thunlp/KB2E ● Evaluators, scorers, baselines and python wrappers of translational models can be found at: https://github.com/D2KLab/entity2rec/tree/dev
- 24. RESULTS System P@5 P@10 MAP NDCG TransH 0.196457 0.170331 0.134170 0.461370 TransR 0.190497 0.165033 0.127401 0.453900 MostPop 0.144603 0.129156 0.092103 0.406294 TransE 0.116656 0.098245 0.071185 0.379548 SVD 0.067815 0.062401 0.042671 0.328776 ItemKNN 0.057483 0.053626 0.040933 0.324887
- 25. CONCLUSIONS ● Knowledge graphs are a beneficial data structure for hybrid recommender systems ● Item recommendation can be seen as a knowledge graph completion problem ● Translational models conceived to predict missing properties in a knowledge graph can be used to provide recommendations ● The empirical comparison shows that TransH outperforms TransR and TransE ● Many possible extensions of this work: considering other datasets, other baselines, other state-of-the-art graph embeddings systems, using knowledge graph embeddings to model multiple user-item interactions
- 26. ● Enrico Palumbo, Data Scientist, ISMB, PhD Student, EURECOM - PoliTo ● Mail: palumbo@ismb.it ● Personal page: http://enricopal.github.io ● Slides: http://www.slideshare.net/EnricoPalumbo2 ● Github: www.github.com/D2KLAB, www.github.com/enricopal ● Twitter: https://twitter.com/enricopalumbo91 Turin, Italy, www.ismb.it Sophia Antipolis, France, www.eurecom.fr Turin, Italy, www.polito.it CONTACTS
- 27. REFERENCES ● [1]: Perozzi, Bryan, Rami Al-Rfou, and Steven Skiena. "Deepwalk: Online learning of social representations." Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2014. ● [2]: Grover, Aditya, and Jure Leskovec. "node2vec: Scalable feature learning for networks." Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2016. ● [3]: Ristoski, Petar, and Heiko Paulheim. "Rdf2vec: Rdf graph embeddings for data mining." International Semantic Web Conference. Springer, Cham, 2016. ● [4]: Palumbo, Enrico, Giuseppe Rizzo, and Raphaël Troncy. "Entity2rec: Learning user-item relatedness from knowledge graphs for top-n item recommendation." Proceedings of the Eleventh ACM Conference on Recommender Systems. ACM, 2017. ● [5]: Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., & Yakhnenko, O. (2013). Translating embeddings for modeling multi-relational data. In Advances in neural information processing systems (pp. 2787-2795). ● [6]: Wang, Z., Zhang, J., Feng, J., & Chen, Z. (2014, July). Knowledge Graph Embedding by Translating on Hyperplanes. In AAAI (Vol. 14, pp. 1112-1119). ● [7]: Lin, Yankai, et al. "Learning entity and relation embeddings for knowledge graph completion." AAAI. Vol. 15. 2015. ● [8]: Harald Steck. 2013. Evaluation of recommendations: rating-prediction and ranking. In Proceedings of the 7th ACM conference on Recommender systems. ACM, 213–220.