Ce diaporama a bien été signalé.
Nous utilisons votre profil LinkedIn et vos données d’activité pour vous proposer des publicités personnalisées et pertinentes. Vous pouvez changer vos préférences de publicités à tout moment.

Max Kuhn's talk on R machine learning

4 740 vues

Publié le

Twitter: @NycDataSci


Learn with our NYC Data Science Program (weekend courses for working professionals and 12 week full time for whom are advancing their career into Data Science)

Our next 12-Week Data Science Bootcamp starts in Jun. (Deadline to apply is May 1st, all decisions will be made by May 15th.)

====================================

Max Kuhn, Director is Nonclinical Statistics of Pfizer and also the author of Applied Predictive Modeling.

He will join us and share his experience with Data Mining with R.

Max is a nonclinical statistician who has been applying predictive models in the diagnostic and pharmaceutical industries for over 15 years. He is the author and maintainer for a number of predictive modeling packages, including: caret, C50, Cubist and AppliedPredictiveModeling. He blogs about the practice of modeling on his website at ttp://appliedpredictivemodeling.com/blog
---------------------------------------------------------

His Feb 18th course can be RSVP at NYC Data Science Academy.

Syllabus
Predictive Modeling using R

Description
This class will get attendees up to speed in predictive modeling using the R programming language. The goal of the course is to understand the general predictive modeling process and how it can be implemented in R. A selection of important models (e.g. tree-based models, support vector machines) will be described in an intuitive manner to illustrate the process of training and evaluating models.

Prerequisites:
Attendees should have a working knowledge of basic R data structures (e.g. data frames, factors etc) and language fundamentals such as functions and subsetting data. Understanding of the content contained in Appendix B sections B1 though B8 of Applied Predictive Modeling (free PDF from publisher [1]) should suffice.


Outline:
- An introduction to predictive modeling

- R and predictive modeling: the good and bad

- Illustrative example

- Measuring performance

- Data splitting and resampling

- Data pre-processing

- Classification trees

- Boosted trees

- Support vector machines

If time allows, the following topics will also be covered
- Parallel processing

- Comparing models

- Feature selection

- Common pitfalls

Materials:
Attendees will be provided with a copy of Applied Predictive Modeling[2] as well as course notes, code and raw data. Participants will be able to reproduce the examples described in the workshop.
Attendees should have a computer with a relatively recent version of R installed.
About the Instructor:
More about Max's work:

[1] http://rd.springer.com/content/pdf/bbm%3A978-1-4614-6849-3%2F1.pdf

[2] http://appliedpredictivemodeling.com

Publié dans : Formation
  • Soyez le premier à commenter

Max Kuhn's talk on R machine learning

  1. 1. An Overview of Predictive Modeling Max Kuhn, Ph.D Pfizer Global R&D Groton, CT max.kuhn@pfizer.com
  2. 2. Outline “Predictive modeling” definition Some example applications A short overview and example How is this different from what statisticians already do? What can drive choice of methodology? Where should we focus our efforts? Max Kuhn Predictive Modeling 2/61 2/61
  3. 3. “Predictive Modeling”
  4. 4. Define That! Rather than saying that method X is a predictive model, I would say: Predictive Modeling is the process of creating a model whose primary goal is to achieve high levels of accuracy. In other words, a situation where we are concerned with making the best possible prediction on an individual data instance. (aka pattern recognition)(aka machine learning) Max Kuhn Predictive Modeling 4/61 4/61
  5. 5. Models So, in theory, a linear or logistic regression model is a predictive model? Yes. As will be emphasized during this talk: the quality of the prediction is the focus model interpretability and inferential capacity are not as important Max Kuhn Predictive Modeling 5/61 5/61
  6. 6. Where Does This Fit Within Data Science From Drew Conway (http://drewconway.com/zia/2013/3/26/the- data-science-venn-diagram) Max Kuhn Predictive Modeling 6/61 6/61
  7. 7. Example Applications
  8. 8. Examples spam detection: we want the most accurate prediction that minimizes false positives and eliminates spam plane delays, travel time, etc. customer volume sentiment analysis of text sale price of a property For example, does anyone care why an email or SMS is labeled as spam? Max Kuhn Predictive Modeling 8/61 8/61
  9. 9. Quantitative Structure Activity Relationships (QSAR) Pharmaceutical companies screen millions of molecules to see if they have good properties, such as: biologically potent safe soluble, permeable, drug–like, etc We synthesize many molecules and run lab tests (“assays”) to estimate the characteristics listed above. When a medicinal chemist designs a new molecule, he/she would like a prediction to help assess whether we should synthesized it. Max Kuhn Predictive Modeling 9/61 9/61
  10. 10. Medical Diagnosis A patient might be concerned about having cancer on the basis of some on physiological condition their doctor has observed. Any result based on imaging or a lab test should, above all, be as accurate as possible (more on this later). If the test does indicate cancer, very few people would want to know if it was due to high levels of: human chorionic gonadotropin (HCG) alpha-fetoprotein (AFP) or lactate dehydrogenase (LDH) Accuracy is the concern Max Kuhn Predictive Modeling 10/61 10/61
  11. 11. Managing Customer Churn/Defection When a customer has the potential to switch vendors (e.g. phone contract ends), we might want to estimate the probability that they will churn the expected monetary loss from churn Based on these quantities, you may want to offer a customer an incentive to stay. Poor predictions will result in loss of revenue from a customer loss or from a inappropriately applied incentive. The goal is to minimize financial loss at the individual customer level Max Kuhn Predictive Modeling 11/61 11/61
  12. 12. An Overview of the Modeling Process
  13. 13. Cell Segmentation Max Kuhn Predictive Modeling 13/61 13/61
  14. 14. Cell Segmentation Individual cell results are aggregated so that decisions can be made about specific compounds Improperly segmented objects might compromise the quality of the data so an algorithmic filter is needed. In this application, we have measurements on the size, intensity, shape or several parts of the cell (e.g. nucleus, cell body, cytoskeleton). Can these measurements be used to predict poorly segmentation using a set of manually labeled cells? Max Kuhn Predictive Modeling 14/61 14/61
  15. 15. The Data The authors scored 2019 cells into these two bins: well–segmented (WS) or poorly–segmented (PS). There are 58 measurements in each cell that can be used as predictors. The data are in the caret package. Max Kuhn Predictive Modeling 15/61 15/61
  16. 16. Model Building Steps Common steps during model building are: estimating model parameters (i.e. training models) determining the values of tuning parameters that cannot be directly calculated from the data calculating the performance of the final model that will generalize to new data Max Kuhn Predictive Modeling 16/61 16/61
  17. 17. Model Building Steps How do we “spend” the data to find an optimal model? We typically split data into training and test data sets: Training Set: these data are used to estimate model parameters and to pick the values of the complexity parameter(s) for the model. Test Set (aka validation set): these data can be used to get an independent assessment of model efficacy. They should not be used during model training. Max Kuhn Predictive Modeling 17/61 17/61
  18. 18. Spending Our Data The more data we spend, the better estimates we’ll get (provided the data is accurate). Given a fixed amount of data, too much spent in training won’t allow us to get a good assessment of predictive performance. We may find a model that fits the training data very well, but is not generalizable (over–fitting) too much spent in testing won’t allow us to get good estimates of model parameters Max Kuhn Predictive Modeling 18/61 18/61
  19. 19. Spending Our Data Statistically, the best course of action would be to use all the data for model building and use statistical methods to get good estimates of error. From a non–statistical perspective, many consumers of of these models emphasize the need for an untouched set of samples the evaluate performance. The authors designated a training set (n = 1009) and a test set (n = 1010). Max Kuhn Predictive Modeling 19/61 19/61
  20. 20. Over–Fitting Over–fitting occurs when a model inappropriately picks up on trends in the training set that do not generalize to new samples. When this occurs, assessments of the model based on the training set can show good performance that does not reproduce in future samples. Some models have specific “knobs” to control over-fitting neighborhood size in nearest neighbor models is an example the number if splits in a tree model Max Kuhn Predictive Modeling 20/61 20/61
  21. 21. Over–Fitting Often, poor choices for these parameters can result in over-fitting For example, the next slide shows a data set with two predictors. We want to be able to produce a line (i.e. decision boundary) that differentiates two classes of data. Two new points are to be predicted. A 5–nearest neighbor model is illustrated. Max Kuhn Predictive Modeling 21/61 21/61
  22. 22. K–Nearest Neighbors Classification Predictor A PredictorB 0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6 q Class 1 Class 2 Max Kuhn Predictive Modeling 22/61 22/61
  23. 23. Over–Fitting On the next slide, two classification boundaries are shown for the a different model type not yet discussed. The difference in the two panels is solely due to different choices in tuning parameters. One over–fits the training data. Max Kuhn Predictive Modeling 23/61 23/61
  24. 24. Two Model Fits Predictor A PredictorB 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 Model #1 0.2 0.4 0.6 0.8 Model #2 Max Kuhn Predictive Modeling 24/61 24/61
  25. 25. Characterizing Over–Fitting Using the Training Set One obvious way to detect over–fitting is to use a test set. However, repeated “looks” at the test set can also lead to over–fitting Resampling the training samples allows us to know when we are making poor choices for the values of these parameters (the test set is not used). Examples are cross–validation (in many varieties) and the bootstrap. These procedures repeated split the training data into subsets used for modeling and performance evaluation. Max Kuhn Predictive Modeling 25/61 25/61
  26. 26. The Big Picture We think that resampling will give us honest estimates of future performance, but there is still the issue of which sub–model to select (e.g. 5 or 10 NN). One algorithm to select sub–models: Define sets of model parameter values to evaluate; for each parameter set do for each resampling iteration do Hold–out specific samples ; Fit the model on the remainder; Predict the hold–out samples; end Calculate the average performance across hold–out predictions end Determine the optimal parameter value; Create final model with entire training set and optimal parameter value;Max Kuhn Predictive Modeling 26/61 26/61
  27. 27. K–Nearest Neighbors Tuning q q q q q q q q q q q q q q q q q q q q q q q q q 0.80 0.81 10 20 30 40 50 #Neighbors Accuracy(RepeatedCross−Validation) Max Kuhn Predictive Modeling 27/61 27/61
  28. 28. Next Steps Normally, we would try different approaches to improving performance: different models, other pre–processing techniques, model ensembles, and/or feature selection (i.e. variable selection) For example, the degree of correlation between the predictors in these data is fairly high. This may have had a negative impact on the model. We can fit another model that uses a sufficient number of principal components in the K–nearest neighbor model (instead of the original predictors). Max Kuhn Predictive Modeling 28/61 28/61
  29. 29. Pre-Processing Comparison q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q 0.80 0.81 0.82 10 20 30 40 50 #Neighbors Accuracy(Cross−Validation) pre_processing q qBasic PCA Max Kuhn Predictive Modeling 29/61 29/61
  30. 30. Evaluating the Final Model Candidate Suppose we evaluated a panel of models and the cross–validation results indicated that the K–nearest neighbor model using PCA had the best cross–validation results. We would use the test set to verify that there were no methodological errors. The test set accuracy was 80.7%. The CV accuracy was 82%. Pretty Close! Max Kuhn Predictive Modeling 30/61 30/61
  31. 31. Comparisons to Traditional Statistical Analyses
  32. 32. Good References On This Topic Breiman, L. (1997). No Bayesians in foxholes. IEEE Expert, 12(6), 21–24. Breiman, L. (2001). Statistical modeling: The two cultures. Statistical Science, 16(3), 199–215. (plus discussants) Ayres, I. (2008). Super Crunchers: Why Thinking-By-Numbers is the New Way To Be Smart, Bantam. Shmueli, G. (2010). To explain or to predict? Statistical Science, 25(3), 289–310. Boulesteix, A.-L., & Schmid, M. (2014). Machine learning versus statistical modeling. Biometrical Journal, 56(4), 588–593. (and other articles in this issue) Max Kuhn Predictive Modeling 32/61 32/61
  33. 33. Empirical Validation Previously there was a emphasis on empirical assessments of accuracy. While accuracy isn’t the best measure, metrics related to errors of some sort are way more important here than in traditional statistics. Statistical criteria (e.g. lack of fit tests, p–values, likelihoods, etc.) are not directly related to performance. Friedman (2001) describes an example related to boosted trees with MLE: “[...] degrading the likelihood by overfitting actually improves misclassification error rates. Although perhaps counterintuitive, this is not a contradiction; likelihood and error rate measure different aspects of fit quality.” Max Kuhn Predictive Modeling 33/61 33/61
  34. 34. Empirical Validation Even some measure that are asymptotically related to error rates (e.g. AIC, BIC) are still insufficient. Also, a number of statistical criteria (e.g. adjusted R2 ) require degrees of freedom. In many predictive models, these do not exist or are much larger than the training set size. Finally, there is often an interest in measuring nonstandard loss functions and optimizing models on this basis. For the customer churn example, the loss of revenue related to false negatives and false positives may be different. The loss function may not fit nicely into standard decision theory so evidence–based assessments are important. Max Kuhn Predictive Modeling 34/61 34/61
  35. 35. Statistical Formalism Traditional statistics are often used for making inferential statements regarding parameters or contrasts of parameters. For this reason, there is a fixation on the appropriateness of the models related to necessary distributional assumptions required for theory relatively simple model structure to keep the math trackable the sanctity of degrees of freedom. This is critical to make appropriate inferences on model parameters. However, letting these go allows the user greater flexibility to increase accuracy. Max Kuhn Predictive Modeling 35/61 35/61
  36. 36. Examples complex or nonlinear pre–processing (e.g. PCA, spatial sign) ensembles of models (boosting, bagging, random forests) overparameterized but highly regularized nonlinear models (SVN, NNets) Quinlan (1993) describing pessimistic pruning, notes that it: “does violence to statistical notions of sampling and confidence limits, so the reasoning should be taken with a grain of salt.” Breiman (1997) notes that there were “No Bayesians in foxholes.” Max Kuhn Predictive Modeling 36/61 36/61
  37. 37. Three Sweeping Generalizations About Models 1 In the absence of any knowledge about the prediction problem, no model can be said to be uniformly better than any other (No Free Lunch theorem) 2 There is an inverse relationship between model accuracy and model interpretability 3 Statistical validity of a model is not always connected to model accuracy Max Kuhn Predictive Modeling 37/61 37/61
  38. 38. Segmentation CV Results for Different Models Confidence Level: 0.95 Accuracy CART LDA glmnet PLS bagging FDA KNN_PCA SVM C5.0 GBM RF 0.79 0.80 0.81 0.82 0.83 0.84 Max Kuhn Predictive Modeling 38/61 38/61
  39. 39. How Does Statistics Contribute to Predictive Modeling? Many statistical principals still apply and have greatly improved the field: resampling the variance–bias tradeoff sampling bias in big data parsimony principal (AOTBE, keep the simpler model) L1 and/or L2 regularization Statistical improvements to boosting are a great example of this. Max Kuhn Predictive Modeling 39/61 39/61
  40. 40. What Can Drive Choice of Methodology?
  41. 41. Cross–Validation Cost Estimates (HPC Case Study, APM, pg 454) Cost C5.0 Random Forests Bagging SVM (Weights) C5.0 (Costs) CART (Costs) Bagging (Costs) SVM Neural Networks CART LDA (Sparse) LDA FDA PLS 0.2 0.3 0.4 0.5 0.6 0.7 0.8 q q q q q q q q q q q q q q q qqq q q q qq q Max Kuhn Predictive Modeling 41/61 41/61
  42. 42. Important Considerations It is fairly common to have a number of different models with equivalent levels of performance. If a predictive model will be deployed, there are a few practical considerations that might drive the choice of model. if the prediction equation is to be numerically optimized, models with smooth functions might be preferable. the characteristics of the data (e.g. multicollinearity, imbalanced, etc) will often constrain the feasible set of models. large amounts of unlabeled data can be exploited by some models Max Kuhn Predictive Modeling 42/61 42/61
  43. 43. Ease of Use/Model Development If a large number of models will be created, using a low maintenance technique is probably a good idea. QSAR is a good example. Most companies maintain a large number of models for different endpoints. These models are constantly updated too. Trees (and their ensembles) tend to be low maintenance; neural networks are not. However, if the model must have the absolute best possible error rate, you might be willing to put a lot of work into the model (e.g. deep neural networks) Max Kuhn Predictive Modeling 43/61 43/61
  44. 44. How will the model be deployed? If the prediction equation needs to be encoded into software or a database, concise equations have the advantage. For example, random forests require a large number of unpruned trees. Even for a reasonable data set, the total forest can have thousands of terminal nodes (39,437 if/then statements for the cell model) Bagged trees require dozens of trees and might yield the same level of performance but with a smaller footprint Some models (K-NN, support vector machines) can require storage of the original training set, which may be infeasible. Neural networks, MARS and a few other models can yield compact nonlinear prediction equations. Max Kuhn Predictive Modeling 44/61 44/61
  45. 45. Ethical Considerations There are some situations where using a sub–optimal model can be unethical. My experiences in molecular diagnostics gave me some perspective on the friction point between pure performance and notions about validity. Another generalization: doctors and regulators are apprehensive about black–box models since they cannot make sense of it and assume that the validation of these models is lacking. I had an experience as a “end user” of a model (via a lab test) when I was developing algorithms for a molecular diagnostic company. Max Kuhn Predictive Modeling 45/61 45/61
  46. 46. Antimicrobial Susceptibility Testing Max Kuhn Predictive Modeling 46/61 46/61
  47. 47. Klebsiella pneumoniae klebsiella-pneumoniae.org/ Max Kuhn Predictive Modeling 47/61 47/61
  48. 48. Where should we focus our efforts?
  49. 49. Diminishing Returns on Models Honestly, I think that we have plenty of effective models. Very few problems are solved by inventing yet another predictive model. Exceptions: using unlabeled data severe class imbalances feature engineering algorithms (e.g. MARS, autoencoders, etc) applicability domain techniques and prediction confidence assessments It is more important to improve the features that go into the model. Max Kuhn Predictive Modeling 50/61 50/61
  50. 50. What About Big Data? The advantages and issues related to Big Data can be broken down into two areas: Big N: more samples or cases Big P: more variables or attributes or fields Mostly, I think the catchphrase is associated more with N than P. Does Big Data solve my problems? Maybe1 1 the basic answer given by every statistician throughout time Max Kuhn Predictive Modeling 51/61 51/61
  51. 51. Can You Be More Specific? It depends on what are you using it for? does it solve some unmet need? does it get in the way? Basically, it comes down to: Bigger Data = Better Data at least not necessarily. Max Kuhn Predictive Modeling 52/61 52/61
  52. 52. Big N - Interpolation One situation where it probably doesn’t help is when samples are added within the mainstream of the data In effect, we are just filling in the predictor space by increasing the granularity. After the first 10,000 or so observations, the model will not change very much. This does pose an interesting interaction within the variance–bias trade off. Big N goes a long way to reducing the model variance. Given this, can high variance/low bias models be improved? Max Kuhn Predictive Modeling 53/61 53/61
  53. 53. High Variance Model with Big N Maybe. Many high(ish) variance/low bias models (e.g. trees, neural networks, etc.) tend to be very complex and computationally demanding. Adding more data allows these models to more accurately reflect the complexity of the data but would require specialized solutions to be feasible. At this point, the best approach is supplanted by the available approaches (not good). Form should still follow function. Max Kuhn Predictive Modeling 54/61 54/61
  54. 54. Model Variance and Bias Predictor Outcome −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 2 4 6 8 10 Max Kuhn Predictive Modeling 55/61 55/61
  55. 55. Low Variance Model with Big N What about low variance/high bias models (e.g. logistic and linear regression)? There is some room here for improvement since the abundance of data allows more opportunities for exploratory data analysis to tease apart the functional forms to lower the bias (i.e. improved feature engineering) or to select features. For example, non–linear terms for logistic regression models can be parametrically formalized based on the results of spline or loess smoothers. Max Kuhn Predictive Modeling 56/61 56/61
  56. 56. Diminishing Returns on N∗ At some point, adding more (∗ of the same) data does not do any good. Performance stabilizes but computational complexity increases. The modeler becomes hand–cuffed to whatever technology is available to handle large amounts of data. Determining what data to use is more important than worrying about the technology required to fit the model. Max Kuhn Predictive Modeling 57/61 57/61
  57. 57. Global Versus Local Models in QSAR When developing a compound, once a “hit” is obtained, the chemists begin to tweak the structure to make improvements. The leads to a chemical series We could build QSAR models on the large number of existing compounds (a global model) or on the series of interest (a local model). Our experience is that local models beat global models the majority of the time. Here, fewer (of the most relevant) compounds are better. Our first inclination is to use all the data because our (overall) error rate should get better. Like politics, All Problems Are Local. Max Kuhn Predictive Modeling 58/61 58/61
  58. 58. Data Quality (QSAR Again) One “Tier 1” screen is an assay for logP (the partition coefficient) which we use as a measure of “greasyness”. Tier 1 means that logP is estimated for most compounds (via model and/or assay) There was an existing, high–throughput assay on a large number of historical compounds. However, the data quality was poor. Several years ago, a new assay was developed that was lower throughput and higher quality. This became the default assay for logP. The model was re-built on a small (N ≈ 1, 000) set chemically diverse compounds. In the end, fewer compounds were assayed but the model performance was much better and costs were lowered. Max Kuhn Predictive Modeling 59/61 59/61
  59. 59. Big N and/or P - Reducing Extrapolation However, Big N might start to sample from rare populations. For example: a customer cluster that is less frequent but has high profitability (via Big N). a specific mutation or polymorphism (i.e. a rare event) that helps derive a new drug target (via Big N and Big P) highly non–linear “activity cliffs” in computational chemistry can be elucidated (via Big N) Now, we have the ability to solve some unmet need. Max Kuhn Predictive Modeling 60/61 60/61
  60. 60. Thanks!

×