1 of 64

Learning Causal Knowledge Graphs from Text Log Data

Aleksandr Eismont

Seminar Novel and Non-mainstream Advances in Data Science (IPD Böhm), 27.01.2022, Karlsruhe

KIT – The Research University in the Helmholtz Association

www.kit.edu

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

2 of 64

Agenda

  • Motivation
  • Approach Structure
  • Log Templates Generation
  • Event Time-Series Creation
  • Causal Inference Algorithms
  • Summary

2

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

3 of 64

Motivation

Causal discovery allows to:

  • Infer causal structure from data
  • Find causes of incidentes
  • Describe effects of events

3

Medicine

​

Economics

Climatology

Cloud System

​

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

4 of 64

Motivation

“It expresses my firm belief that the current data-fitting direction taken by “Data Science” is temporary (read my lips!), that the future of “Data Science” lies in causal data interpretation…”

4

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

5 of 64

Approach Structure

5

Raw Log Data

Event Time-Series

Log Templates

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

6 of 64

Approach Structure

6

Raw Log Data

Event Time-Series

Log Templates

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

7 of 64

Log Templates Generation

State-of-the-art algorithms:

  1. Drain [1]
  2. IPLoM [2]
  3. NuLog [3]

7

  • original log message

1998 05-30-2016 10:02:30.146 sshd[1234]: Accepted password for ops from 192.0.2.3 port 12345

​

  • log template

sshd[***]: Accepted password for *** from *** port ***

[S. Kobayashi et al., “Mining Causality of Network Events in Log Data”, 2018]

Raw Log Data

Log Templates

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

8 of 64

Approach Structure

8

Raw Log Data

Event Time-Series

Log Templates

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

9 of 64

Event Time-Series Creation

9

template_2

[T. Jia et al., “An Approach for Anomaly Diagnosis based on Hybrid Graph Model with Logs for Distributed Services”, 2017]

time

template_1

9

time

Log Templates

Event Time-Series

Frequency Computing

template_2

time

template_1

time

. . . . .

. . . . . .

.

. . . . .

. . . . . . .

.

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

10 of 64

Approach Structure

10

Raw Log Data

Event Time-Series

Log Templates

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

11 of 64

Causal Inference Algorithm

  • Constraint-based
    • PC Algorithm
  • Score-based
    • Greedy Equivalence Search Algorithm
  • Functional-based
    • Linear Non-Gaussian Acyclic Model
    • Post-nonlinear Model
  • Granger Causality

​

11

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

12 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm
    1. Causal Assumptions
    2. Conditional Independence

​

12

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

13 of 64

Causal Inference Algorithm

Constraint-based

  • Conditional Independence
    • G-square test

    • Fisher-Z test

​

​

    • Conditional Mutual Information

​

​

13

[R.E. Neapolitan, “Learning Bayesian Networks”, 2004]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

14 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm

​

​

14

Events

[P. Spirtes et al., “Causation, Prediction, and Search”, 2001]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

15 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm

​

​

15

Events

Complete graph

[P. Spirtes et al., “Causation, Prediction, and Search”, 2001]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

16 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm

​

​

16

Events

Complete graph

Remove edges between independent events

[P. Spirtes et al., “Causation, Prediction, and Search”, 2001]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

17 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm

​

​

17

Events

Complete graph

Remove edges between independent events

Apply V-structure

[P. Spirtes et al., “Causation, Prediction, and Search”, 2001]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

18 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm

​

​

18

Events

Complete graph

Remove edges between independent events

Apply V-structure

Apply orientation rules

[P. Spirtes et al., “Causation, Prediction, and Search”, 2001]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

19 of 64

Causal Inference Algorithm

Constraint-based

  • PC Algorithm variations:
    • PC-stable [4]
      • order-independent
      • longer processing time
    • Parallel-PC [5]
      • parallelisation in the conditional independence tests
    • PC-simple [6]
      • target variable
    • HITON-PC [7]
      • order-independent
      • target variable
    • MPC [8]
      • extra orientational rule against cycles

19

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

20 of 64

Causal Inference Algorithm

Score-based

  • Greedy Equivalence Search (GES) Algorithm
    • Greedy principle
    • Bayesian Information Criterion (BIC)

​

​

20

[D.M. Chickering, “Optimal Structure Identification With Greedy Search”, 2002]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

21 of 64

Causal Inference Algorithm

Score-based

  • Greedy Equivalence Search (GES) Algorithm

​

21

Events

[D.M. Chickering, “Optimal Structure Identification With Greedy Search”, 2002]

Forward Equivalence Search (FES)

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

22 of 64

Causal Inference Algorithm

Score-based

  • Greedy Equivalence Search (GES) Algorithm

​

22

Events

FSE Local Maximum

[D.M. Chickering, “Optimal Structure Identification With Greedy Search”, 2002]

Forward Equivalence Search (FES)

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

23 of 64

Causal Inference Algorithm

Score-based

  • Greedy Equivalence Search (GES) Algorithm

​

23

Events

FSE Local Maximum

[D.M. Chickering, “Optimal Structure Identification With Greedy Search”, 2002]

Forward Equivalence Search (FES)

Backward Equivalence Search (BES)

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

24 of 64

Causal Inference Algorithm

Score-based

  • Greedy Equivalence Search (GES) Algorithm

​

24

Events

FSE Local Maximum

BES Local Maximum

[D.M. Chickering, “Optimal Structure Identification With Greedy Search”, 2002]

Forward Equivalence Search (FES)

Event Time-Series

Causal Knowledge Graph

Backward Equivalence Search (BES)

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

25 of 64

Causal Inference Algorithm

Score-based

  • Greedy Equivalence Search (GES) Algorithm variations:
    • Greedy Interventional Equivalence Search (GIES) [9]
      • Turning Step
    • Fast Greedy Equivalence Search (FGES) [10]
      • parallelisation
      • causal faithfulness assumption

​

​

25

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

26 of 64

Causal Inference Algorithm

Functional-based

  • Data asymmetry

​

26

[S. Shimizu et al., “DirectLiNGAM: A Direct Method for Learning a Linear Non-Gaussian Structural Equation Model”, 2011]

X

Y

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

27 of 64

Causal Inference Algorithm

Functional-based

  • Data asymmetry

​

27

[S. Shimizu et al., “DirectLiNGAM: A Direct Method for Learning a Linear Non-Gaussian Structural Equation Model”, 2011]

X

Y

Event Time-Series

Causal Knowledge Graph

?

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

28 of 64

Causal Inference Algorithm

Functional-based

  • Data asymmetry

​

28

[S. Shimizu et al., “DirectLiNGAM: A Direct Method for Learning a Linear Non-Gaussian Structural Equation Model”, 2011]

Event Time-Series

Causal Knowledge Graph

X

Y

?

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

29 of 64

Causal Inference Algorithm

Functional-based

  • Data asymmetry

​

29

[S. Shimizu et al., “DirectLiNGAM: A Direct Method for Learning a Linear Non-Gaussian Structural Equation Model”, 2011]

Event Time-Series

Causal Knowledge Graph

X

Y

?

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

30 of 64

Causal Inference Algorithm

Functional-based

  • Data asymmetry

​

30

[S. Shimizu et al., “DirectLiNGAM: A Direct Method for Learning a Linear Non-Gaussian Structural Equation Model”, 2011]

Event Time-Series

Causal Knowledge Graph

X

Y

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

31 of 64

Causal Inference Algorithm

Functional-based

  • Linear Non-Gaussian Acyclic Model (LiNGAM) [11]
    • Function is linear
    • at most one of the noise term E and cause X is Gaussian
  • Post-nonlinear (PNL) Causal Model [12]
    • Functions are nonlinear

​

31

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

32 of 64

Causal Inference Algorithm

Functional-based

​

​

32

Events

Complete graph

Remove edges between independent events

[S. Kobayashi et al., “A Quantitative Causal Analysis for Network Log Data”, 2021]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

33 of 64

Causal Inference Algorithm

Functional-based

​

​

33

Events

Complete graph

Remove edges between independent events

Apply LiNGAM

[S. Kobayashi et al., “A Quantitative Causal Analysis for Network Log Data”, 2021]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

34 of 64

Causal Inference Algorithm

Granger Causality

  • The cause occurs before the effect
  • The cause contains unique information about the effect that is not available otherwise

34

[C.W.J. Granger, “Investigating Causal Relations by Econometric Models and Cross-spectral Methods”, 1969]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

35 of 64

Approach Structure

35

Raw Log Data

Event Time-Series

Log Templates

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

36 of 64

Summary

  • Log Template Generation
    • Clustering
    • NLP
  • Event Time-Series Creation
    • Sliding Window
  • Causal Inference Algorithm
    • Constraint-based
    • Score-based
    • Functional-based
    • Granger Causality

36

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

37 of 64

References

[1] He, P., Zhu, J., Zheng, Z., Lyu, M.R.: “Drain: An Online Log Parsing Approach with Fixed Depth Tree”, 2017

[2] Makanju, A., Nur Zincir-Heywood, A., Milios, E.: “A Lightweight Algorithm for Message Type Extraction in System Application Logs”, 2012

[3] Nedelkoski, S., Bogatinovski, J., Acker, A., Cardoso, J., Kao, O.: “Self-Supervised Log Parsing”, 2020

[4] Colombo, D., Maathuis, M.H.: “Order-Independent Constraint-Based Causal Structure Learning”, 2012

[5] Le, T.D., Hoang, T., Li, J., Liu, L., Liu, H., Hu, S.: “A Fast PC Algorithm for High Dimensional Causal Discovery with Multi-Core PCs”, 2015

[6] Bühlmann, P., Kalisch, M., Maathuis, M.H.: “Variable selection in high-dimensional linear models: partially faithful distributions and the PC-simple algorithm”, 2009

[7] Aliferis, C.F., Tsamardinos, I., Statnikov, A.: “HITON: A Novel Markov Blanket Algorithm for Optimal Variable Selection”, 2003

[8] Tsagris, M.: “Bayesian Network Learning with the PC Algorithm: An Improved and Correct Variation”, 2018

[9] Hauser, A., Bühlmann, P.: “Characterization and Greedy Learning of Interventional Markov Equivalence Classes of Directed Acyclic Graphs”, 2011

[10] Ramsey, J., Glymour, M., Sanchez-Romero, R., Glymour, C.: “A million variables and

more: The Fast Greedy Equivalence Search algorithm for learning high-dimensional graphical causal models, with an application to functional magnetic resonance images”, 2017

[11] Shimizu, S., Hoyer, P., Hyvärinen, A., Kerminen, A.. “A Linear Non-Gaussian Acyclic Model for Causal Discovery”, 2006

[12] Zhang, K., Hyvärinen, A.: “On the Identifiability of the Post-Nonlinear Causal Model”, 2009

37

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

38 of 64

Log Templates Generation

  • Clustering
    1. LogMine [1]
    2. LKE [2]
  • Frequent pattern mining
    • SLCT [3]
    • LFA [4]
  • Evolutionary
    • MoLFI [5]
  • Log-structure heuristics
    • Drain [6]
    • AEL [7]
  • Longest-common sub-sequence
    • Spell [8]
  • Neural
    • NuLog [9]

38

  • original log message

1998 05-30-2016 10:02:30.146 sshd[1234]: Accepted password for ops from 192.0.2.3 port 12345

​

  • log template

sshd[***]: Accepted password for *** from *** port ***

[S. Nedelkoski et al., “Self-Supervised Log Parsing”, 2020]

Raw Log Data

Log Templates

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

39 of 64

Event Time-Series Creation

39

time

template_1

39

[T. Jia et al., “An Approach for Anomaly Diagnosis based on Hybrid Graph Model with Logs for Distributed Services”, 2017]

raw log

raw log

raw log

raw log

log_template_1

log_template_2

log_template_3

log frequency

time

template_2

time

template_3

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

40 of 64

Event Time-Series Creation

  1. Sliding Window

40

time

template_1

40

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

41 of 64

Event Time-Series Creation

  • Sliding Window

41

time

template_1

41

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

42 of 64

Event Time-Series Creation

  • Sliding Window

42

time

template_1

42

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

43 of 64

Event Time-Series Creation

  • Sliding Window

​

​

​

​

  • Thresholding

43

time

template_1

43

time

template_1

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

44 of 64

Event Time-Series Creation

  • Sliding Window

​

​

​

​

  • Thresholding

44

time

template_1

44

time

template_1

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

45 of 64

Event Time-Series Creation

  • Sliding Window

​

​

​

​

  • Thresholding

45

time

template_1

45

time

template_1

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

46 of 64

Event Time-Series Creation

  • Sliding Window

​

​

​

​

  • Thresholding

46

time

template_1

46

time

template_1

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

47 of 64

Event Time-Series Creation

a) Sliding window

b) Thresholding

47

time

template_1

47

time

template_1

[C. Glymour et al., “Review of Causal Discovery Methods Based on Graphical Models”, 2019]

Log Templates

Event Time-Series

t

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

48 of 64

Causal Inference Algorithm

Constraint-based

  • Assumptions:
    1. Causal Markov – if two variables are conditional independent of each other, given all their intermediate variables, these variables must be unconnected
    2. Causal Faithfulness – if two variables are conditional dependent, there must be an edge connecting those variables
    3. Causal Sufficiency – all the common causes of a pair of variables are measured (there are no hidden/external causes)
    4. Independency – no variable is an (indirect) cause of itself

​

48

[A.R. Nogueira et al., “Causal Discovery in Machine Learning: Theories and Applications”, 2021]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

49 of 64

Causal Inference Algorithm

Constraint-based

  • Conditional Independence

​

49

Height

Vocabulary

Child

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

50 of 64

Causal Inference Algorithm

Constraint-based

  • Conditional Independence

​

50

Age

Height

Vocabulary

Child

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

51 of 64

Causal Inference Algorithm

Constraint-based

  • Conditional Independence

​

51

Age

Height

Vocabulary

Child

Height and Vocabulary of Child are

conditional independent given Age of Child

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

52 of 64

Causal Inference Algorithm

Constraint-based

  • Conditional Independence

​

52

Age

Height

Vocabulary

Child

Height and Vocabulary of Child are

conditional independent given Age of Child

First throw

Second throw

Dice

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

53 of 64

Causal Inference Algorithm

Constraint-based

  • Conditional Independence

​

53

Age

Height

Vocabulary

Child

Height and Vocabulary of Child are

conditional independent given Age of Child

Sum

First throw

Second throw

Dice

The First and the Second throws of Dice are conditional dependent given the Sum of them

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

54 of 64

Causal Inference Algorithm

Granger Causality

​

​

54

[M. Eichler, “Causal Inference in Time Series Analysis”, 2012]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

55 of 64

Causal Inference Algorithm

Granger Causality

​

​

  • Null Hypothesis (H0): Xt does not Granger-cause Yt
  • F-test

55

[M. Eichler, “Causal Inference in Time Series Analysis”, 2012]

Event Time-Series

Causal Knowledge Graph

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

56 of 64

Causal Inference Algorithm

56

Events

FSE Local Maximum

BES Local Maximum

[D.M. Chickering, “Optimal Structure Identification With Greedy Search”, 2002]

Forward Equivalence Search (FES)

Backward Equivalence Search (BES)

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

57 of 64

This talk is about...

asdasdasd

  • asdasd
  • asdasd

​

asdasdasd

  • asd
    • asdasd
  • asdasd
    • asdasd
    • asdasd

​

57

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

58 of 64

Example diagram

58

Insights

Predictions

Recommendations

Data

Machine Learning Model

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

59 of 64

Example diagram

59

Engineering and �Environmental Systems

Healthcare

Other methods �with implicit domain knowledge

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

60 of 64

Use info boxes to deliver main messages

60

Engineering and �Environmental Systems

It is a very heterogeneous field with 400+ use case specific studies.

Recent surveys structure the field by grouping various methods into taxonomies.

Up to date there are no existing energy-related publications.

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

61 of 64

Example Workflow

61

Step 1

Step 2

Step 3

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

62 of 64

This talk is about...

Repeat and remind the agenda slide when you switch the sections

  • asdasd
  • asdasd

​

asdasd

  • Before we asdasd
    • asdasd
  • Now we asdasd
    • asdasd

​

62

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

63 of 64

Summary

asdasd

  • asdasd
  • asdasd

​

63

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)

64 of 64

Backup

64

Aleksandr Eismont: Learning Causal Knowledge Graphs from Text Log Data

Seminar Novel and Non-mainstream Advances

in Data Science (IPD Böhm)