Outflow: Exploring Flow, Factors and Outcome of Temporal Event Sequences
1 like1,999 views
The document presents an exploration of event sequences and their outcomes, particularly in the context of sports and other fields. It details a multi-step layout process for visualizing these sequences, including aggregation, visual encoding, and graph drawing techniques. The results from a user study suggest the tool 'Outflow' effectively supports event sequence analysis tasks.
![Assumption
• Events are persistent.
Entity #1
e1 e2 e3
Entity #1
e1 e1 e1
[e1] e2 e2
e3
States [e1, e2]
[e1, e2, e3]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-21-320.jpg)
![Select alignment point
Pick a state
What are the paths What are the paths
that led to ? after ?
or just an empty state []
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-23-320.jpg)
![Outflow Graph
Alignment Point
[e1, e2, e3]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-24-320.jpg)
![1 entity
Outflow Graph
Alignment Point
[e1]
[e1, e2]
[ ]
[e1, e2, e3]
[e1, e2, e3, e5]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-25-320.jpg)
![2 entities
Outflow Graph
Alignment Point
[e1]
[e1, e2]
[ ]
[e1, e3]
[e1, e2, e3]
[e1, e2, e3, e5]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-26-320.jpg)
![3 entities
Outflow Graph
Alignment Point
[e1]
[e1, e2]
[e1, e2, e3, e4]
[ ]
[e1, e3]
[e1, e2, e3]
[e1, e2, e3, e5]
[e3]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-27-320.jpg)
![n entities
Outflow Graph
Alignment Point
[e1]
[e1, e2]
[e1, e2, e3, e4]
[ ]
[e2]
[e1, e3]
[e1, e2, e3]
[e1, e2, e3, e5]
[e3]
[e2, e3]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-28-320.jpg)
![n entities
Outflow Graph
Alignment Point
[e1]
[e1, e2]
[e1, e2, e3, e4]
[ ]
[e2]
[e1, e3]
[e1, e2, e3]
[e1, e2, e3, e5]
[e3]
[e2, e3]
Average outcome = 0.4
Average time = 10 days
layer Number of entities = 10
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-29-320.jpg)
![Factors
Time
Entity #1
[e1] [e1, e2] [e1, e2, e3]
Factor 1 Factor 2 Factor 3 Factor 4
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-55-320.jpg)
![Factors
Time
Patient #1
[e1] [e1, e2] [e1, e2, e3]
Yellow Injury Red Substitution
Which factors are correlated to each state?
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-56-320.jpg)
![Present factors
Alignment Point
Factor 1 [e1] [e1,e2]
[e1,e2,e3,e4]
[] [e2] [e1,e3]
[e1,e2,e3]
[e1,e2,e3,e5]
[e3] [e2,e3]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-58-320.jpg)
![Absent factors
Alignment Point
[e1] [e1,e2]
[e1,e2,e3,e4]
Factor 2
[] [e2] [e1,e3]
Factor 2
[e1,e2,e3]
[e1,e2,e3,e5]
[e3] [e2,e3]
m](https://image.slidesharecdn.com/outflowtalkfinalshare-121018172424-phpapp01/85/Outflow-Exploring-Flow-Factors-and-Outcome-of-Temporal-Event-Sequences-59-320.jpg)
Outflow: Exploring Flow, Factors and Outcome of Temporal Event Sequences
- 1. InfoVis 2012 Seattle, WA Outflow Exploring Flow, Factors and Outcomes of Temporal Event Sequences Krist Wongsuphasawat HCIL, University of Maryland David Gotz IBM Research m
- 2. m Events
- 3. m Event | 12:15 p.m. Lunch
- 4. m Event Sequences Event Event Event
- 5. Daily Activity 7:30 a.m. 7:45 a.m. 8:15 a.m. Wake Up Exercise Go to work m
- 6. Soccer Game 10th minute 25th minute 90th minute Team A scores Team B scores Team A scores m
- 7. Soccer Game Time Game #1 10th minute 25th minute 90th minute Goal Concede Goal m
- 8. Many games Time Game #1 Goal Concede Goal Game #2 Goal Goal Concede Game #3 Goal Concede Concede Game #n Concede Goal Goal Goal m
- 9. with outcome Time Game #1 Win (1) Goal Concede Goal Game #2 Win (1) Goal Goal Concede Game #3 Lose (0) Goal Concede Concede Game #n Win (1) Concede Goal Goal Goal m
- 10. 7 events per entity 7 event types 823543 co mbinations m
- 11. Enjoy! m
- 12. consumable m
- 13. Overview / Summary Event Sequences with Outcome m
- 14. m 7 Steps
- 15. m Step 1 | Aggregation
- 16. Event Sequences Entity #1 Entity #2 Outflow Graph Entity #3 Entity #4 Entity #5 Entity #6 Entity #7 … Entity #n m
- 17. Assumption • Events are persistent. Entity #1 e1 e2 e3 Entity #1 m
- 18. Assumption • Events are persistent. Entity #1 e1 e2 e3 Entity #1 e1 e1 e1 m
- 19. Assumption • Events are persistent. Entity #1 e1 e2 e3 Entity #1 e1 e1 e1 e2 e2 m
- 20. Assumption • Events are persistent. Entity #1 e1 e2 e3 Entity #1 e1 e1 e1 e2 e2 e3 m
- 21. Assumption • Events are persistent. Entity #1 e1 e2 e3 Entity #1 e1 e1 e1 [e1] e2 e2 e3 States [e1, e2] [e1, e2, e3] m
- 22. Select alignment point Pick a state What are the paths What are the paths that led to ? after ? Example Soccer: Goal, Concede, Goal m
- 23. Select alignment point Pick a state What are the paths What are the paths that led to ? after ? or just an empty state [] m
- 24. Outflow Graph Alignment Point [e1, e2, e3] m
- 25. 1 entity Outflow Graph Alignment Point [e1] [e1, e2] [ ] [e1, e2, e3] [e1, e2, e3, e5] m
- 26. 2 entities Outflow Graph Alignment Point [e1] [e1, e2] [ ] [e1, e3] [e1, e2, e3] [e1, e2, e3, e5] m
- 27. 3 entities Outflow Graph Alignment Point [e1] [e1, e2] [e1, e2, e3, e4] [ ] [e1, e3] [e1, e2, e3] [e1, e2, e3, e5] [e3] m
- 28. n entities Outflow Graph Alignment Point [e1] [e1, e2] [e1, e2, e3, e4] [ ] [e2] [e1, e3] [e1, e2, e3] [e1, e2, e3, e5] [e3] [e2, e3] m
- 29. n entities Outflow Graph Alignment Point [e1] [e1, e2] [e1, e2, e3, e4] [ ] [e2] [e1, e3] [e1, e2, e3] [e1, e2, e3, e5] [e3] [e2, e3] Average outcome = 0.4 Average time = 10 days layer Number of entities = 10 m
- 30. Soccer Results Alignment Point 1-0 2-0 2-2 0-0 1-1 2-1 3-1 0-1 0-2 m
- 31. m Step 2 | Visual Encoding
- 32. Past Future Alignment Node’s horizontal position shows sequence of states. e1! e2! e3! End of path e1! e1! e2! time link e1! Node’s height is edge edge e2! number of entities. e4! e2! Color is outcome Time edge’s width is measure. duration of transition. m
- 33. m Step 3 | Graph Drawing
- 34. m
- 35. m
- 36. 3.1 Sugiyama’s heuristics • Directed Acyclic Graph (DAG) layout – Sugiyama, K., Tagawa, S. & Toda, M., 1981. Methods for Visual Understanding of Hierarchical System Structures. IEEE Transactions on Systems, Man, and Cybernetics, 11(2), p.109-125. • Reduce edge crossing m
- 37. 41 crossings m
- 38. 12 crossings m
- 39. m
- 40. 3.2 Force-directed layout • Spring simulation Each node is particle. x Total force = Force from edges - Repulsion between nodes m
- 41. m
- 42. m
- 43. 3.3 Edge Routing • Avoid unnecessary crossings Reroute m
- 44. 3.3 Edge Routing • After routing m
- 45. m
- 46. m
- 47. m Step 4 | Interactions
- 48. Interactions • Panning • Zooming • Brushing • Pinning • Tooltip • Event type selection m
- 49. m Demo
- 50. m Step 5 | Simplification
- 51. Node Clustering • Cluster nodes in each layer • Similarity measure: Outcome, etc. • Threshold (0-1) m
- 52. m
- 53. m
- 54. m Step 6 | Factors
- 55. Factors Time Entity #1 [e1] [e1, e2] [e1, e2, e3] Factor 1 Factor 2 Factor 3 Factor 4 m
- 56. Factors Time Patient #1 [e1] [e1, e2] [e1, e2, e3] Yellow Injury Red Substitution Which factors are correlated to each state? m
- 57. Information Retrieval Which keywords are correlated to each document? State 1 State 2 State 3 … … … Factor xxx … … … … … Doc#1 Doc#2 Doc#3 Which factors are correlated to each state? m
- 58. Present factors Alignment Point Factor 1 [e1] [e1,e2] [e1,e2,e3,e4] [] [e2] [e1,e3] [e1,e2,e3] [e1,e2,e3,e5] [e3] [e2,e3] m
- 59. Absent factors Alignment Point [e1] [e1,e2] [e1,e2,e3,e4] Factor 2 [] [e2] [e1,e3] Factor 2 [e1,e2,e3] [e1,e2,e3,e5] [e3] [e2,e3] m
- 60. tf-idf • Term frequency tf = Number of times a term t appear in the document Number of terms in the document • Inverse document frequency idf = log ( Number of documents Number of documents that has the term t + 1 ) m
- 61. Score based on tf-idf • Ratio (presence) Rp = Number of entities with factor f before state Number or entities in the state • Inverse state ratio (presence) R-1 sp = log ( Number of states Number of states preceded by factor f + 1 ) m
- 62. m
- 63. m Step 7 | User Study
- 64. User Study • Goal: Evaluate Outflow’s ability to support event sequence analysis tasks • 12 participants • 60 minutes each • 9 tasks + 7 training tasks • Questionnaire m
- 65. Results • Accurate: 3 mistakes from 108 tasks • Fast: Average 5-60 seconds • Findings: – From video – Different outcomes for each incoming paths – Etc. m
- 66. Future Work • Integration with prediction algorithm • Additional layout techniques • Advanced factor analysis • Deeper evaluations with domain experts m
- 67. Conclusions • Event sequences with outcome • Outflow – Interactive visual summary – Explore flow & outcome – Factors – Multi-step layout process • Not specific to sports Contact: kristw@twitter.com dgotz@us.ibm.com @kristwongz m
- 68. Heart failure (CHF) patient Time Patient #1 Die (0) Aug 1998 Oct 1998 Jan 1999 Ankle Edema Cardiomegaly Weight Loss m
- 69. Event Sequences Medical Transportation Sports Education Web logs Logistics and more… m
- 70. Acknowledgement • Charalambos (Harry) Stavropoulos • Robert Sorrentino • Jimeng Sun • Comments from HCIL colleagues m
- 71. Conclusions • Event sequences with outcome • Outflow – Interactive visual summary – Explore flow & outcome – Factors – Multi-step layout process • Not specific to medical or sports Contact: kristw@twitter.com dgotz@us.ibm.com @kristwongz m