GoConqr has a new look! Let us know what you think.

Sign up free

Flashcards

MGTS3206

joel_clement34 English 44 cards

MGTS3206

Cards in this deck (44)

  1. 1MGTS3206 Lecture 1 Data Scientistdo actual modelling
    statistician
  2. 2Modelrepresentation of reality
  3. 3Predictive model Prediction InstancePM - formula estimating unknown value of interest
    P - estimate an unknown value
    I - represents a fact or data point
  4. 4Classification example Regression Similarity Matching Clustering Data ReductionC - how likely is consumer to respond
    R - how much will she use the service
    SM - can find consumers similar?
    C - do my customers form groups
    DR - which latent dimensions describe taste preference
  5. 5Co-occurrence Grouping Profiling Link Prediction Causal modelingCG- association rule, basket analysis
    P - normal behavior look like
    LP - John and Jane share friends should they become friends
    Causal Modeling - why are consumers leaving
  6. 6Supervised Methodsclassification
    regression
    causal modeling
    similarity matching
    link prediction
    data reduction
  7. 7Classification RegressionC - categorical target often binary (class probability estimation)
    R - numeric target
  8. 8“Will this customer purchase service“How much will this customer use the service?”
    Regression problem
    - Numeric target
    - Target variable: amount of usage per customer
  9. 9Summary of bad habitsmissing all the application
    sick people came from a very artificial subset
    target is not target
    no way of telling how the model will perform or test it
  10. 10Expected Value of InformationFirst how to handle continuous probability distribution
    Second, what happen when there are many uncertain events and information is available about some or all of them
    Evaluate decision option with and without the information, and find the difference in the EMV
  11. 11Value of informationsensitivity analysis, structuring
    1. tornado diagram
    2. Perform sensitivity analysis
    3. Probabilistic model
  12. 12Splitting criterioninformation gain (purity measure called entropy)
  13. 13Tree structure rulesNo two parents share descendants
    no cycles
    branches always down
    always end up at a lead node with class determination
  14. 14Classification problemtarget takes on discrete values that are NOT ordered
    binary classification where target is either 0 or 1
  15. 15Solutions to classificationclassifier model
    Binary: Ranking, Probability estimation
  16. 16PruningPost-pruning (fully grown cuts unreliable parts)
    Pre pruning
    Post pruning is preferred
  17. 17Decision pruningDecision pruning
  18. 18Classification functionIf 1 * Age - 1.5 * Balance + 60 > 0
  19. 19Best line of fitobjective function should represent our goal
  20. 20Distinction between classification and regression is if target variable is categorical or numeric (Logistic regression produces numeric estimates (values in the target variable are categorical)Logistic regression is a class probability estimation model not a regression model
  21. 21SVM Support vector machinesUSe hinge loss - example that is not on the wrong side of the margin (hinge loss only positive when an example is on the wrong side)
  22. 22Zero-one loss Squared errorZOL - assigns loss of zero for correct decision and one for incorrect decision
    SE - specifies a loss proportional to the square of the distance from the boundary
  23. 23Ranking Class Probability EstimationR - Tree induction
    R - Linear discriminant functions (Linear regressions, Logistic regressions, SVM)
    CPE - tree induction, Logistic regression
  24. 24Linear model vs Tree InductionSmooth modelling
    non-linear
    how much data
  25. 25Panel Data analysisimportance of fixed and random effects
    Hausman test (fixed and random effects)
  26. 26Panel datacross sectional unit and time series
  27. 27Panel data use- explicit account of individual specific heterogeneity
    - more data variation
    - dynamics of change
    0 detecting and measuring
    study of more complex behavioral models
  28. 28Fixed effect and random effectLarge and small N likely to be little difference so FE
    Large n and small t estimates differ significantly (cross-sectional groups are random) RE
    if error component v correlates with x then RE is biased and FE is not
    For Large N and small T and assumptions behind RE
  29. 29Hausman Testtest difference between coefficient estimates obtained by FE and RE then null hypothesis that the RE estimates are efficient and consistent and FE estimates are inefficient
  30. 30Lecture 5 Model Performance AnalysisOverfitting data (chance patterns that do not generalize is over-fitting)
  31. 31Over-fittingmodel memorizes the properties of the particular training set rather than learning the underlying concept or phenomenon
  32. 32Holdout validationGeneralization - holdout set for final evaluation is test set
  33. 33Holdout evaluation to cross validationstatistics on estimated performance
    better use of a limited data set (cross validation computes its estimates over all the data)
  34. 34Generalization performancemodelling procedures difference
    different training sets
    different test sets
    different test sets different estimates
    training set size
  35. 35Logistic regression vs Tree inductionsmaller training set = logistic regression
    classification trees are more flexible model
    flexibility of tree induction can be advantage with larger training sets
  36. 36Learning curve show generalization performance plotted against amount of training data usedfitting graph shows the generalization performance as well as the performance on the training data but plotted against model complexity
  37. 37Avoiding over-fittingpost pruning
    pre pruning
    Linear models - feature selection, regularization (optimize some combination of fit and simplicity)
  38. 38Nearal Networks for Data Mining (applicationstarget customer
    bank loan
    hiring
    stock purchase
    trading electricity
  39. 39ANNANN
  40. 40Backpropagationlearning alogrithm - comparing computed outputs to desired outputs of historical cases
    Network Structure - (input intermediate (hidden) (output)
  41. 41Loan Processing Example Inputs – income level, age, home ownership nodes Outputs – approval/disapproval of loan nodesConnection weights
    The weight associated with each link in a neural network model
    strength of the data transferred between layers in the network
    They are assessed by neural networks learning algorithms
  42. 42Threshold value is sometimes used instead of a transformation function A hurdle value for the output of a neuron to trigger the next level of neurons. If an output value is smaller than the threshold value, it will not be passed to the next level of neuronsData collection and preparation
    The data used for training and testing must include all the attributes that are useful for solving the problem
    Recall the bankruptcy prediction problem we modeled using logistic regression -- The same data can be used to train a neural network:
    working capital/total assets (WC/TA)
    retained earnings/total assets (RE/TA)
    earnings before interest and taxes/total assets (EBIT/TA)
    market value of equity/total debt (MVE/TD)
    sales/total assets (S/TA)
  43. 43Implementing an ANNrequires interfaces with other IS
    monitoring and feedback
    confidence of users and management in early deployment
  44. 44—

Information

Description
Flashcards on MGTS3206, created by joel_clement34 on 14/04/2016.
Created by
joel_clement34 · Apr 2016
Updated
Apr 22, 2016
Views
0
Rating

Feedback

What is it about?

What do you think of the new look?

How do you like it?