XAI-lecture 2

• Human data interaction is based on

  • understanding patterns
  • gain insights
  • make decision
  • communicate
• Why visualization --> humans can see patterns that alogrithms cannot
• Visualization 


  Computer-based visualization systems provide visual representations of datasets designed to help people carry out tasks more efficiently
• Human is needed in the loop if the system is ill-defined or ill-structured (no single optimal solution, no clear objective measures)

  if fully automated solution exist + trusted --> not needed
• we need visual representations


  perception beats cognition
• data_ink ratio = (data ink)/ total ink used in graphic
• Bigger datasets risk of creating hairball

  solution :

  1. interaction
  2. combing alogrithms and vis
• combining algorithms and vis : explorator visual anaylics <--> interactive visualisation
• Explainability, interpretability,

  intelligibility, and transparency often

  used interchangeably
• '"Inmates Running the Asylum Problem"

  • XAI used to be done by ML researches
  • Algorithms-centered XAI often criticized
  • XAI solutions developed often based on researchers intuition only

  goal :

  • Design XAI solutions for needs of their intended audience
  • Consider how various users interpret and react to explanations
• WHO is the human in XAI?

  cater to diverse types of users and stakeholders
• Performance vs. explainability is the tradeoff for ML techniques
• ML 4 Vis usage example --> help humans finding the good views
• Vis excursion storytelling :

  • introduction
  • problem
  • climax
  • resolution
  • conclusion

  A) Problem

  B) Introduction

  C) Climax

  D) Resolution

  E) Conclusion

  
• Good stories do more than provide facts and data, they situate and give context, they engage, they educate
• Author driven :

  • linear ordering
  • heavy messaging
  • no interactivity
• Reader driven :

  • no ordering
  • no messaging
  • free interactivity
• martini glass structure 


  start with author driven, open up for exploration
• interactive slideshow


  split into multiple scenes allow interaction mid-way
• drill-down story : 


  let reader decide which path to follow, all paths are annotated
• Bump chart: A chart that shows the frequency of occurrence of a particular event, time progress from left to right, brightness indicates the value
• Grouping principles :

  • proximity
  • containment (common region)
  • connection
  • similarity (colors f.e)
  • continuation
  • common fate
• Grouping principles

  A) Proximity

  B) connection

  C) similarity

  D) continuation

  E) common fate

  F) containment

  
• grouping principles (strong -> to less)

  connection is a very strong grouping principle

  A) proximity

  B) color

  C)

  D) size

  E) Shape

  
• Bumps charts sorting algortihms

  A) Quickshort

  B) Bubblesort

  C) selection sort

  D) heapsort

  E) insertion nsort

  F) shell short

  
• KMeans algortihm :

  1. choose the number of k clusters
  2. select random K points as clusters centes
  3. assign each data point to the nearest centroid
  4. compute and place the new centroid of each cluster
  5. repeat step 4 until no observations change cluster
• Bayes' theorem

  • A,B = events
  • P(A|B) = probability of A given B is true
  • P(A), P(B) = the independant probabilities of A and B

  $P(A|B) =$$\frac{P(B|A) P(A)}{P(B)}$
• sensitivity : 


  true positive rate, recall rate
• sensitivity


  proportion of positives which are correctly identified as such
• specifity

  true negative rate
• specificity

  proportion of negatives which are correctly identified as such
• sensitivity example : percentage of sick people who are correctly identifieda s having the condition

  specficity example : percentage of healthy people who are correctly identified as not having the condition
• Data mining --> find correct answer

  Visualization --> understand answer