![Basic Cypher Queries
MATCH p=(t:Talk)-[r:IS_GIVEN_BY]->(s:Speaker)
WHERE s.name= “Geoffrey” AND s.surname= “Hinton” RETURN p](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-24-320.jpg)
![Basic Cypher Queries
MATCH p=(g:Guest)-[r:ATTENDS]->(t:Talk)
WHERE s.name= “A” AND s.surname= “Hinton” RETURN p](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-25-320.jpg)
![Basic Cypher Queries
MATCH p=(s:Speaker)<-[:IS_GIVEN_BY]-(t:Talk)<-[:ATTENDS]-(g:Guest)
WHERE s.name="Geoffrey" and s.surname="Hinton"
RETURN p](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-26-320.jpg)
![ML - 101
X
Input
=Layer 0
Output
1ST
Layer=Z[1]
Z ŷ
y = mx + b](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-36-320.jpg)
![ML - 101 - Forward Propagation
x1
x2
Input
=Layer 0
Output
1ST
Layer=Z[1]
Z1
ŷ
Z2
y = w1
x1
+ w2
x2
+ b](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-37-320.jpg)
![ML - 101 Backward Propagation
x1
x2
Input
=Layer 0
Output
1ST
Layer=Z[1]
Z1
ŷ
Z2
y = w1
x1
+ w2
x2
+ b](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-38-320.jpg)
![ML - 101 Model & Prediction
x1
x2
Input
=Layer 0
Output
1ST
Layer=Z[1]
Z1
Z2
ŷ
X =
x1 # of Beds
x2 PostCode
y = House sales price](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-39-320.jpg)
![Neural Networks - 101
x1
x2
Input
=Layer 0
Output
1ST
Layer=Z[1]
X =
z1
1
z1
2
z1
3
a1
1
a1
2
a1
3
a2
z2
ŷ
a3
z3
a2
z2
z1
4
2ND
Layer=Z[2]
3RD
Layer=Z[3]
X1
X2
a1
4](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-40-320.jpg)
![Neural Networks - 101
x1
x2
Input
=Layer 0
Output
1ST
Layer=Z[1]
X =
z1
1
z1
2
z1
3
a1
1
a1
2
a1
3
a2
z2
ŷ
a3
z3
a2
z2
a
2
z2
2ND
Layer=Z[2]
3RD
Layer=Z[3]
X1
X2
z= ∑Wi
xi
+ b
z= ∑Wx1
+ ∑Wx2
+ b
X =
a= σ(Z)
Z[1]
=
z11
z12
z21
z22
z31
z32
z41
z42
X1
X2](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-41-320.jpg)
![Layer 0 Output
1 Layer=Z[1]
a2
ŷ
Challenge](https://image.slidesharecdn.com/cdwtalk-220118093855/85/Evolution-of-Graph-Algorithms-Benefits-and-Challenges-42-320.jpg)
Evolution of Graph Algorithms – Benefits and Challenges
- 1. Abstract Graphs are the most native format to represent the data, it has built-in structure to store the contextual information and enable data analysts to get better answers for their queries, data engineers to easily extend their eco-system, and data scientists to build enhanced models. We can make informed decisions. In this presentation, you will learn about the evolution of the Graphs with a review on knowledge graphs, path finding, clustering, and recommendation algorithms, how they are built and their use cases. Also, we will review the challenges we face as processing more data comes with a cost, and how they are solved in different faces of the machine learning lifecycle.
- 2. Evolution of Graph Algorithms – Benefits and Challenges By Ebru Cucen @ebrucucen 2/12/2021
- 3. Throughout this talk, I will show graphs in two representations with - Left Hand Side : the data - Right Hand Side : the schema
- 7. Goal Graphs are the most native data format Graphs are useful for all roles in the data habitat Graphs opened a new era in ML
- 9. Historical Timeline Euler’s Koninsberg bridge “Graph” introduced PageRank Google’s Knowledge Graph Semantic Web Facebook’s Entities Graph Yahoo’s Knowledge Graph Gori, GNN McCulloch-Pitts Neural Nets DeepMind Kipf,GCN 1878 1735 1998 1999 2013 2012 2005 2016 2010 1943
- 10. Interest in Research *Title and Abstract search at Dimensions.ai “Graph” and “Neural” and “Network” * “Graph” and “Network” *
- 13. Roles Data Analysts Data/ML Engineers Data Scientists
- 17. Let’s ask
- 20. https://cacm.acm.org/magazines/2019/8/238342-industry-scale-knowledge-graphs/fulltext Users - Ebay - Product recommendation: node and link similarity - Microsoft - Bing Knowledge Graph: Q&A - Academic Graph - LinkedIn Graph - IBM - Watson Discovery - Extend core Knowledge Graph - Facebook - Social Network Analysis
- 21. World Avatar https://www.sciencedirect.com/science/article/abs/pii/S0098135419301589
- 24. Basic Cypher Queries MATCH p=(t:Talk)-[r:IS_GIVEN_BY]->(s:Speaker) WHERE s.name= “Geoffrey” AND s.surname= “Hinton” RETURN p
- 25. Basic Cypher Queries MATCH p=(g:Guest)-[r:ATTENDS]->(t:Talk) WHERE s.name= “A” AND s.surname= “Hinton” RETURN p
- 26. Basic Cypher Queries MATCH p=(s:Speaker)<-[:IS_GIVEN_BY]-(t:Talk)<-[:ATTENDS]-(g:Guest) WHERE s.name="Geoffrey" and s.surname="Hinton" RETURN p
- 28. E2E Process
- 30. What is the shortest path from Victoria to Liverpool Street Station?
- 32. Inherently Graph Use cases ● Telephone network ● Internet packet routing ● Manufacturing
- 34. Inherently Graph Use cases ● Telephone network ● Internet packet routing ● Manufacturing ● Human Brain ● Neural nets are networks
- 35. # Graph Neural Networks
- 36. ML - 101 X Input =Layer 0 Output 1ST Layer=Z[1] Z ŷ y = mx + b
- 37. ML - 101 - Forward Propagation x1 x2 Input =Layer 0 Output 1ST Layer=Z[1] Z1 ŷ Z2 y = w1 x1 + w2 x2 + b
- 38. ML - 101 Backward Propagation x1 x2 Input =Layer 0 Output 1ST Layer=Z[1] Z1 ŷ Z2 y = w1 x1 + w2 x2 + b
- 39. ML - 101 Model & Prediction x1 x2 Input =Layer 0 Output 1ST Layer=Z[1] Z1 Z2 ŷ X = x1 # of Beds x2 PostCode y = House sales price
- 40. Neural Networks - 101 x1 x2 Input =Layer 0 Output 1ST Layer=Z[1] X = z1 1 z1 2 z1 3 a1 1 a1 2 a1 3 a2 z2 ŷ a3 z3 a2 z2 z1 4 2ND Layer=Z[2] 3RD Layer=Z[3] X1 X2 a1 4
- 41. Neural Networks - 101 x1 x2 Input =Layer 0 Output 1ST Layer=Z[1] X = z1 1 z1 2 z1 3 a1 1 a1 2 a1 3 a2 z2 ŷ a3 z3 a2 z2 a 2 z2 2ND Layer=Z[2] 3RD Layer=Z[3] X1 X2 z= ∑Wi xi + b z= ∑Wx1 + ∑Wx2 + b X = a= σ(Z) Z[1] = z11 z12 z21 z22 z31 z32 z41 z42 X1 X2
- 42. Layer 0 Output 1 Layer=Z[1] a2 ŷ Challenge
- 44. Karate Club
- 45. Karate Club - Adjacency List - Adjacency Matrix N1 N2 N3 N1 N2 N3
- 50. Torch.Geometric GCNConv Karate club with GCN
- 51. Final Output Karate club with GCN Put original graph
- 52. Calculating Loss Function Node/Edge/Graph classification Zi =f(hi ) Zij =f(hi ,hj ,eij ) ZG =f(∑i hi)
- 53. Calculating Loss Function Node/Edge/Graph classification Zi =f(hi ) Zij =f(hi ,hj ,eij ) ZG =f(∑i hi)
- 54. Random Walks - Drunk person walking - Equal probability of paths to adjacent nodes - Max uncertainty (local entropy) - Measure structure of graph - Sampling
- 55. # Summary
- 56. Conclusion • Graph is the most native data format with built-in semantic • Graph opened a new era in machine learning none of us can ignore • Graph is very useful for all roles for E2E process
- 57. # Questions