High Profile 🔝 8250077686 📞 Call Girls Service in GTB Nagar🍑
A Review of Relational Machine Learning(SRL) for Knowledge Graphs
1. A Review of Relational Machine Learning
for Knowledge Graphs
Present By: Yalda Akbarzadeh
May 2018
2. A Review of
Relational Machine Learning
for Knowledge Graphs
1. INTRODUCTION
2. KNOWLEDGE
GRAPHS
A. Knowledge
representation
B. Open vs.
closed world
assumption
C. Knowledge
base
construction
3. STATISTICAL
RELATIONAL LEARNING
FOR KNOWLEDGE
GRAPHS
D. Uses of
knowledge
graphs
E. Main tasks in
knowledge graph
construction and
curation
A. Probabilistic
knowledge
graphs
B. Statistical
properties of
knowledge
graphs
C. Types of
SRL models
4. LATENT
FEATURE MODELS
D. Multi-layer
perceptrons
B. Other tensor
factorization
models
C. Matrix
factorization
methods
A. RESCAL: A
bilinear model
E. Neural tensor
networks
F. Latent
distance models
G. Comparison
of models
8. MARKOV
RANDOM
FIELDS
C. Path Ranking
Algorithm
B. Rule Mining and
Inductive Logic
Programming
A. Similarity
measures for uni-
relational data
6. COMBINING
LATENT AND GRAPH
FEATURE MODELS
7. TRAINING SRL
MODELS ON
KNOWLEDGE
GRAPHS 5. GRAPH FEATURE
MODELS
A. Additive
relational
effects model
B. Other
combined
models
A. Penalized
maximum
likelihood
training
B. Where do
the negative
examples come
from?
C. Pairwise loss
training
D. Model
selection
B. Inference
A. Representation
C. Learning
D. Discussion
9. KNOWLEDGE
VAULT
10. EXTENSIONS
AND FUTURE
WORK
D. Discussion
B. Hard
constraints
A. Non-binary
relations
C. Generalizing
to new entities
and relations
D. Querying
probabilistic
knowledge graphs
12. APPENDIX
11. CONCLUDING
REMARKS
D. Querying
probabilistic
knowledge graphs
3. A Review of Relational Machine Learning for Knowledge Graphs
1. INTRODUCTION
2. STATISTICAL RELATIONAL LEARNING FOR KNOWLEDGE GRAPHS
3. LATENT FEATURE MODELS
4. GRAPH FEATURE MODELS
5. COMBINING LATENT AND GRAPH FEATURE MODELS
6. TRAINING SRL MODELS ON KNOWLEDGE GRAPHS
7. MARKOV RANDOM FIELDS
EXTENSIONS AND FUTURE WORK, COCLUSION
REFERENCES
4. Traditional machine
Learning Algorithms
feature vector,
numeric or categorical attribute
Input
Mapping
Output Observation
tower
Class labels,
Regression score,
Unsupervised cluster id
Latent vector (embedding)
1. INTRODUCTION A. Relational Machine Learning 4
5. feature vector,
numeric or categorical attribute
Representation of an Object
Traditional machine
Learning Algorithms
Graph,
Node and Labeled edges
Berlin
Munich
Düsseldorf
478 Km.
486 Km.
505Km.
1. INTRODUCTION A. Relational Machine Learning 5
6. Some Canonical Tasks
>> Collective classification
>> Linkprediction
>> Link-based clustering
prediction of
missing edge
clustering nodes
based on their
connectivity pattern
prediction of
properties of
nodes
1. INTRODUCTION A. Relational Machine Learning 6
9. >> Knowledge graphs model information in the form of entities and relationships between them.
Linked Data
(subject, predicate, object) :: (SPO)
(Leonard Nimoy, played, Spock)
Leonard Nimoy was an actor who played the character
Spock in the science-fiction movie Star Trek
Semantic Web
Linked Data
Resource Description Framework
subject object
predicate
1. INTRODUCTION B. KNOWLEDGE GRAPHS i. Knowledge representation 9
10. >> closed world assumption (CWA),
non-existing triples indicate false relationships.
Linked Data>> KGs are known to be very incomplete
>> RDF and the Semantic Web make the open-world assumption.
>> open world assumption (OWA),
a non-existing triple is interpreted as unknown.
1. INTRODUCTION B. KNOWLEDGE GRAPHS ii. Open vs. closed world assumption 10
Existing triples are facts
What about non-existing
triples?
13. Linked Data
1. INTRODUCTION B. KNOWLEDGE GRAPHS v. Main tasks in KGs construction and curation
Link prediction Entity resolution Link-based clustering
Kate
Ted
Jack
is sister of
isfatherof
isfatherof
Jack
Bonn
uni.
Worksin
uni.
Of
Bonn
Ted
Studiesin
TjLu
Rj
Ed
Lj
Rt
Kj
Jd
Pj
13
16. 2. STATISTICAL RELATIONAL LEARNING FOR KNOWLEDGE GRAPHS B. Statistical properties of KGs 16
• Denotes that query answer
should follow a type.
• Tendency of entities to be related
to other entities with similar characteristic.
• All the same blocks members have
the same relation with other block.
17. 2. STATISTICAL RELATIONAL LEARNING FOR KNOWLEDGE GRAPHS C. Types of SRL models 17
M1) latent features and additional parameters
M2) observed graph features and additional parameters
M3) existence of local interactions (Markov Random Fields)
19. 3. LATENT FEATURE MODELS 19
prestigious awardlatent features -->
Ex) Guinness received the Academy Award
good actor
<- good actor ->
<- prestigious award ->
Relationships between entities can be derived from interactions of their latent features.
20. 3. LATENT FEATURE MODELS A. RESCAL 20
>> RESCAL is a bilinear relational latent feature model which explains triples via pairwise interactions of latent
features.
This importance is used here:
Relationships between entities can be derived
from interactions of their latent features.
21. 3. LATENT FEATURE MODELS B. Matrix factorization methods 21
>> Approach for learning from knowledge graphs is based on matrix factorization
memory
complexity
Lose
information
Reshape
24. 3. LATENT FEATURE MODELS C. Multi-layer-perceptrons 24
It can reduce the number of required parametersIt can reduce the number of required parameters
Advantage of E-MLP
Disadvantage of E-MLP
02
03
04
Σ
26. 3. LATENT FEATURE MODELS E. Latent distance models 26
02
03
04
>> Latent distance models or latent space models in social network analysis:
Derive the probability of relationships from the distance between latent representations of entities
#--------------------------------------------------------------------------------------------------------------------------------------
27. 4. GRAPH FEATURE MODELS A. Similarity measures for uni-relational data 27
>> Existence of an edge can be predicted by extracting features from the observed edges in the graph.
>> GFMs Widely used for link prediction in uni-relational graphs.
>> intuition:
similar entities are likely to be related (homophily)
the similarity of entities can be derived from:
• neighborhood of nodes
• existence of paths between nodes
>> Approaches to measure the similarity of entities:
John Mary
Parent of
Anne
Parent of
(John, marriedTo, Mary)
Predict
Local similarity
Global similarity
Quasi-local similarity
28. 4. GRAPH FEATURE MODELS B. Path Ranking Algorithm 28
>> Using random walks of bounded lengths for predicting links
in multi-relational knowledge graphs.
29. 4. GRAPH FEATURE MODELS B. Path Ranking Algorithm 29
EXAMPLES OF PATHS LEARNED BY PRA ON FREEBASE
TO PREDICT WHICH COLLEGE A PERSON ATTENDED
Relation Path Weight
(draftedBy, school) 2.62
(sibling(s), sibling, education, institution) 1.88
(parents, education, institution) 1.37
(children, education, institution) 1.85
>> key idea in PRA:
Use path probabilities as features for predicting the probability of missing edges.
30. 5. COMBINING LATENT AND GRAPH FEATURE MODEL 30
>>Latent feature models:
well-suited for modeling global relational patterns via newly
introduced latent variables.
They are computationally efficient if triples can be explained
with a small number of latent variables.
>>Graph feature models:
are well-suited for modeling local and quasi-local graphs
patterns.
They are computationally efficient if triples can be explained
from the neighborhood of entities or from short paths in
the graph
31. 5. COMBINING LATENT AND GRAPH FEATURE MODEL Additive Relational Effects models 31
32. 6. TRAINING SRL MODELS ON KNOWLEDGE GRAPHS A. Penalized maximum likelihood training 32
02
03
04
posterior priorlikelihood
33. 6. TRAINING SRL MODELS ON KNOWLEDGE GRAPHS B. Where do the negative examples come from? 33
>> Most knowledge graphs only contain positive training examples.
36. 7. MARKOV RANDOM FIELDS 36
02
03
04
>> Form of Markov Random Fields (MRFs):
37. EXTENSIONS AND FUTURE WORK 37
A. Non-binary relations
B. Hard constraints: types, functional constraints, and others
C. Generalizing to new entities and relations
D. Querying probabilistic knowledge graphs
E. Trustworthiness of knowledge graphs
38. Conclusion 38
review of state-of-the-art statistical relational learning (SRL) methods applied to very large knowledge graphs
KGs have found important applications in question answering, structured search, exploratory search
showed how to create a truly massive, machine-interpretable “semantic memory” of facts
39. REFERENCES
39
[1] M. Nickel, K. Murphy, V. Tresp, and E. Gabrilovich, “A review of relational machine learning for knowledge
graphs,” Proceedings of the IEEE, vol. 104, pp. 11-33, 2016.
[2] N. Lao, T. Mitchell, and W. W. Cohen, “Random walk inference and learning in a large scale knowledge base,”
in Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2011, pp. 529-539.
[3] “An Introduction to Statistical Relational Learning” at: www.cs.kuleuven.be/~lucdr/salvador.pdf.
[4] “sampling procedure” at: https://www.slideshare.net/anasomoray/chapter-4-sampling-procedure
[5] “Non-Hierarchical Clustering” at: https://courses.cs.washington.edu/courses/cse527/01au/oct25/oct25.html
[6] https://en.wikipedia.org/
[7] N. Feldman, Learning and overgeneralization patterns in a connectionist model of the German plural: Cite seer,
2005.
[8] “Introduction to: Open World Assumption vs Closed World Assumption”, at:
http://www.dataversity.net/introduction-to-open-world-assumption-vs-closed-world-assumption/
[9] “Introducing the Knowledge Graph ”, at: https://www.youtube.com/watch?v=mmQl6VGvX-c
[10] “What is Linked Data?”, at: https://www.w3.org/standards/semanticweb/data
[11] “Resource Description Framework (RDF)”, at: https://www.w3.org/TR/rdf-concepts/