1 / 121

Turning Bayesian Model Averaging Into Bayesian Model Combination

Turning Bayesian Model Averaging Into Bayesian Model Combination. Kristine Monteith , James L. Carroll, Kevin Seppi , Tony Martinez Presented by James L. Carroll At LANL CNLS 2011, and AMS 2011 LA-UR 11-05664. Abstract.

jock
Download Presentation

Turning Bayesian Model Averaging Into Bayesian Model Combination

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Turning Bayesian Model Averaging Into Bayesian Model Combination Kristine Monteith, James L. Carroll, Kevin Seppi, Tony Martinez Presented by James L. Carroll At LANL CNLS 2011, and AMS 2011 LA-UR 11-05664

  2. Abstract Bayesian methods are theoretically optimal in many situations. Bayesian model averaging is generally considered the standard model for creating ensembles of learners using Bayesian methods, but this technique is often outperformed by more ad hoc methods in empirical studies. The reason for this failure has important theoretical implications for our understanding of why ensembles work. It has been proposed that Bayesian model averaging struggles in practice because it accounts for uncertainty about which model is correct but still operates under the assumption that only one of them is. In order to more effectively access the benefits inherent in ensembles, Bayesian strategies should therefore be directed more towards model combination rather than the model selection implicit in Bayesian model averaging. This work provides empirical verification for this hypothesis using several different Bayesian model combination approaches tested on a wide variety of classification problems. We show that even the most simplistic of Bayesian model combination strategies outperforms the traditional ad hoc techniques of bagging and boosting, as well as outperforming BMA over a wide variety of cases. This suggests that the power of ensembles does not come from their ability to account for model uncertainty, but instead comes from the changes in representational and preferential bias inherent in the process of combining several different models.

  3. Supervised UBDTM . .. DTest X F X . .. D Y Y DTrain O End Use Utility

  4. The Mathematics of Learning • Learning about F:

  5. The Mathematics of Learning • Learning about F: Repeat to get

  6. The Mathematics of Learning • Learning about F: • Classification or Regression:

  7. The Mathematics of Learning • Learning about F: • Classification or Regression: • Decision making:

  8. The Mathematics of Learning • Learning about F: • Classification or Regression: • 0-1 Loss Decision making:

  9. Realistic Machine Learning: Training Using Training Data Input: Unlabeled Instance Learning Algorithm hypothesis hypothesis Output: Class label

  10. Using a Learner • Info about • sepal length • sepal width • petal length • petal width hypothesis Iris Setosa Iris Virginica Iris Versicolor

  11. Using a Learner • Info about • sepal length • sepal width • petal length • petal width hypothesis Iris Setosa Iris Virginica Iris Versicolor

  12. Using a Learner • Info about • sepal length • sepal width • petal length • petal width hypothesis Can we do better? Iris Setosa Iris Virginica Iris Versicolor

  13. Ensembles • Multiple learners vs.

  14. Creating Ensemble Diversity: data1 hypothesis1 data2 hypothesis2 Learning Algorithm Training Data data3 hypothesis3 data4 hypothesis4 data5 hypothesis5

  15. Creating Ensemble Diversity : algorithm1 hypothesis1 algorithm2 hypothesis2 Training Data algorithm3 hypothesis3 algorithm4 hypothesis4 algorithm5 hypothesis5

  16. CLASSIFYING AN INSTANCE: Input: Unlabeled Instance h1 h2 h3 h4 h5 Iris Setosa: 0.1 Iris Virginica: 0.3 Iris Versicolor: 0.6 Iris Setosa: 0.3 Iris Virginica: 0.3 Iris Versicolor: 0.4 Iris Setosa: 0.4 Iris Virginica: 0.5 Iris Versicolor: 0.1 Iris Setosa: Iris Virginica: Output: Class label

  17. CLASSIFYING AN INSTANCE: Input: Unlabeled Instance h1 h2 h3 h4 h5 Iris Setosa: 0.1 Iris Virginica: 0.3 Iris Versicolor: 0.6 Iris Setosa: 0.3 Iris Virginica: 0.3 Iris Versicolor: 0.4 Iris Setosa: 0.4 Iris Virginica: 0.5 Iris Versicolor: 0.1 Iris Setosa: Iris Virginica: Output: Class label

  18. Possible OPTIONS FOR COMBINING HYPOTHESES: • Bagging: One hypothesis, One vote • Boosting: Weight by predictive accuracy on the training set • BAYESIAN MODEL AVERAGING (BMA): Weight by the formal probability that each hypothesis is correct given all the data xi: Unlabeled Instance yi: Probability of class label

  19. BMA Two Steps: • Step 0 • Train learners • Step 1 • Grade learners • Step 2 • Use learners

  20. BMA Two Steps: • Step 0 • Train learners • Step 1 • Grade learners • Step 2 • Use learners • Optimal Solution:

  21. BMA Two Steps: • Step 0 • Train learners • Step 1 • Grade learners • Step 2 • Use learners

  22. BAYESIAN TECHNIQUES “Bayes is right, and everything else is wrong, or is a (potentially useful) approximation.” -James Carroll Please compare your algorithm to Bayesian Model Averaging - Reviewer for a conference where Kristine submitted her thesis research on ensemble learning

  23. BMA is the “Optimal” Ensemble Technique? “Given the ‘correct’ model space and prior distribution, Bayesian model averaging is the optimal method for making predictions; in other words, no other approach can consistently achieve lower error rates than it does.” - Pedro Domingos

  24. DOMINGOs’ Experiments • Domingos decided to put this theory to the test. • 2000 empirical study of ensemble methods: • J48 • Bagging • BMA

  25. DOMINGOs’ Experiments

  26. DOMINGOs’ Experiments

  27. DOMINGOs’ Experiments

  28. Domingos’s Observation: • Bayesian Model Averaging gives too much weight to the “maximum likelihood” hypothesis

  29. Domingos’s Observation: • Bayesian Model Averaging gives too much weight to the “maximum likelihood” hypothesis Compare two classifiers with 100 data points: One with 95% predictive accuracy and one with 94% predictive accuracy Bayesian Model Averaging weights the first classifier as 17 TIMES more likely!

  30. Clarke’s Experiments • 2003, comparison between BMA and stacking. • Similar results to Domingos • BMA is vastly outperformed by stacking

  31. Clarke’s Claim: h1 h2 h3

  32. Clarke’s Claim: DGM h1 h2 h3

  33. Clarke’s Claim: DGM h1 Projection h2 h3

  34. Clarke’s Claim: h1 DGM h2 h3

  35. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  36. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  37. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  38. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  39. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  40. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  41. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  42. Clarke’s Claim: • BMA converges to model closest to the Data Generating Model (DGM) instead of converging to the combination closest to DGM! h1 DGM h2 h3

  43. Is Clark Correct?

  44. Is Clark Correct?

  45. Is Clark Correct?

  46. Is Clark Correct?

  47. Is Clark Correct? • 5 samples

  48. Is Clark Correct? • 10 samples

  49. Is Clark Correct? • 15 samples

  50. Is Clark Correct? • 20 samples

More Related