Reliable AI

Question 1

data shift (concept shift/drift, fracture point)

Answer 1

• when the training and the test distributions are different
• lose confidence in the ability to predict
• causes:
  
  1. sample selection bias
  2. non-stationary (spatial or temporal) shifts in data
• formula: P_test(y, x) != P_train(y, x)

Question 2

covariate shift

Answer 2

• change in covariate (independent variable) distributions
• causes: caused by
  
  • temporal, spatial changes
  • data sparsity
  • biased feature selection
  • class shift
• formula: P_test(y|x) = P_train(y|x) and P_train(x)≠P_test(x)

Question 3

prior probability shift

Answer 3

• changes in the distribution of the class
• causes:
  
  1. class imbalance
• formula: P_test(x|y) = P_train(x|y) and P_train(y)≠P_test(y)

Question 4

concept shift

Answer 4

• related to the relationship between the input and output distribution
• e.g. before and after financial crises needs a concept shift
• formula:
  
  • P_train(y|x)≠P_test(y|x) and P_train(x)=P_test(x), for X→Y
  • P_train(x|y)≠P_test(x|y) and P_train(y)=P_test(y), for Y→X

Question 5

internal covariate shift

Answer 5

• variation shift of the activations
• batch normalization should hasten learning (also regularizes the input by adding noise)

Question 6

statistical similarity

Answer 6

• non-parametric: kl divergence, jensen-shannon divergence, population stability index, wasserstein distance
• multi-variate: for high dimensional and unstructured datasets
  
  • isolation forest
  • KD Trees
  • variational autoencoder
  • normalization flow

Question 7

statistical distance

Answer 7

• compare histograms of training data over time
  
  • population stability index:
  • kolmogorov-smirnov statistic:
  • Kullback-Lebler Divergence:
• problems:
  
  • not good for high-dimensional
  • not good for sparse features

Question 8

novelty detection

Answer 8

• create a model for modeling distribution
  
  • one-class support vector machine
• good: good for complex interactions
• bad: cannot tell you explicitly what has changed

Question 9

discriminative distance

Answer 9

• less common method
• train classifier to detect whether point is from the source domain
• use training error as proxy of distance between distributions (high error means closer)
• pros:
  
  • may be only feasible solution for some deep learning methods (e.g. NLP)
  • good for sparse data
  • good for high dimensions
• problems:
  
  • can only be done offline
  • more complicated than other methods

Question 10

how to handle dataset shift

Answer 10

1. remove features: set a boundary of acceptable shift, remove and retrain if too much shift
   
   1. remove if it is no important
2. reweigh importance: upweight training data more similar to test instance
3. adversarial search: uses adversarial model that is robust to feature deletion