2. Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019 2
WHAT ARE WE TALKING ABOUT?
3. Recommender Systems
3
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
4. Graph-based Data Model
Very popular representation
for hybrid recommender
systems
Recommendations are
obtained by running
algorithms such as
PageRank and Spreading
Activation
Very good performance in
recommendation tasks [*]
4
[*] Cataldo Musto, Pierpaolo Basile, Pasquale Lops,
Marco de Gemmis, Giovanni Semeraro:
Introducing linked open data in graph-based
recommender systems. Inf. Process. Manage. 53(2):
405-435 (2017)
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
Users = nodes Items = nodes
Characteristics of the items = nodes from
Preferences & Properties= edges
5. Intuition: Graph Embedding techniques
5
Graph embedding
techniques take a
graph as input and
build a vector-space
representation of the
nodes (and the
relations, eventually)
as output.
They resulted as very effective in several machine learning tasks.
What about a recommender system based on graph embedding techniques?
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
6. Research Questions
6
How effective is a
hybrid
recommendation
method based on
graph embedding
techniques?
How do features
extracted from
DBpedia impact on
the overall
effectiveness of the
representation?
1. 2.
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
7. 7
METHODOLOGY
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
8. Methodology
8
Step 1 – build a graph-based data model. Two alternatives: bipartite
(without DBpedia features ) or tripartite (with DBpedia features)
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
9. Methodology
9
Step 2 – run a graph embedding technique over the data model,
and build a vector space representation for each user and item
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
10. Methodology
10
Step 3 – use the vectors to feed a classification algorithm in a node classification
task, that labels the nodes (items) as positive or negative for the user
?
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
11. 11
EXPERIMENTS
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
12. Experimental Design
12
Datasets: ML1M -
Librarythings – Last.fm
Data Model:
collaborative, DBpedia-
based, complete
DBpedia features: all
Methodology - Step 1
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
13. Experimental Design
13
Graph Embeddings
Techniques: Laplacian
Eigenmaps (a community-
preserving technique) and
Node2Vec (a structural-
preserving technique)
Size of the Vectors: 128,
256, 512 Methodology - Step 2
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
14. (Recap of Graph Embedding techniques)
14
Community-preserving technique
(e.g., Laplacian Eigenmaps)
Structural-preserving technique
(e.g., Node2Vec)
Input Output
Source: Goyal, Palash, and Emilio Ferrara. "Graph
embedding techniques, applications, and
performance: A survey." Knowledge-Based Systems
151 (2018): 78-94.
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
15. Experimental Design
15
Classification
Algorithm: Logistic
Regression
Metrics: F1-measure
Methodology - Step 3
?
3 datasets x 2 techniques x 3 embedding size
x 3 data models = 54 runs
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
16. Results
16
1. Node2Vec > Laplacian Eigenmaps
2. Complete (tripartite) Graph > Collaborative (bipartite) Graph
3. Tiny gaps by increasing the size of the vectors
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
17. Results
Our
framework
overcomes all
the baselines
0,6886
0,6559
0,6117
0,597 0,5935
0,5115
0,5835
0,5935
0,5092
0,5297
0,5961
0,5117
0,6023 0,6034
0,599
0,6032
0,5964
0,5642
0,6083
0,6152 0,6107
0,5
0,55
0,6
0,65
0,7
MovieLens Librarything Last.fm
GE+LOD U2U-KNN I2I-KNN BPRMF PR BPRMF+LOD PPR+LOD
U2U-KNN: User-to-User
Collaborative Filtering
I2I-KNN: Item-to-Item
Collaborative Filtering
BPRMF: Bayesian Personalized
Ranking Matrix Factorization
PPR: Personalized PageRank
Advanced baselinesSimple baselines DBpedia-based baselines
17Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
18. Take-home Messages
1.
2.
Graph embedding techniques can be an effective alternative to
classic graph-based data models. Structural-preserving techniques
tend to obtain better results.
18
Exogenous descriptive features extracted from knowledge graphs
as DBpedia further improve the accuracy of the model. The size of
the vectors does not significantly affects the overall effectiveness
of the recommendation methodology
3. Improvement over state-of-art recommendation methods!
Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019
19. Cataldo Musto, Pierpaolo Basile, Giovanni Semeraro. Hybrid Semantics-aware Recommendation Exploiting Knowledge Graph Embeddings.
AI*IA 2019, 18th International Conference of the Italian Association for Artificial Intelligence, Rende, November 19, 2019 19
Grazie!
cataldo.musto@uniba.it
@cataldomusto
Contacts