際際滷

際際滷Share a Scribd company logo

Caret and zoon: machine learning, ecology and domain specific package systems

?
?
Tim Lucas
Tim Lucas

Talk on two R packages, caret and zoon. Presenting to the Oxford rusers group. I'll discuss the scope and uses of both packages. I'll also discuss whether domain specific package ecosystems (which zoon aims to be) are a good idea and what makes them useful, difficult and different.

1 of 29
Download to read offline
Caret and zoon: machine learning, ecology and domain speci?c package systems Tim C.D. Lucas Malaria Atlas Project, BDI, Oxford @timcdlucas @statsforbios timcdlucas@gmail.com
Who am I? Malaria Atlas Project at BDI Malaria, maps, geostatistics
Who am I? R packages Zoon INLAutils palettetown - my greatest ever achievement
Talk overview caret General package for machine learning. Introduction to the package. A domain speci?c package ecosystem? zoon General package for species distribution modelling. What are SDMs? Package overview. Domain speci?c ecosystems Other examples. Are they a good thing?
caret https://topepo.github.io/caret/model-training-and-tuning.html
What is machine learning? q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q Not cat Cat 0.00 0.25 0.50 0.75 Feature / predictor variable Class 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 2 4 6 0.25 0.50 0.75 1.00 Feature / predictor variable Response
Cross-validation
Hyperparameters
Hyperparameters Number of PCA coordinates Cut-o?s for variable selection x + x2 + x3 + x4 + ...
No free lunch No such thing as a universal, ｀best¨ machine learning model.
What does caret do?
What does caret do?
Training a model m1 <- train(Species ~ ., iris, method = ｀gbm¨) q q q 0.935 0.940 0.945 50 75 100 125 150 # Boosting Iterations Accuracy(Bootstrap) Max Tree Depth q 1 2 3
Training a di?erent model m2 <- train(Species ~ ., iris, method = ｀nnet¨) q q q 0.75 0.80 0.85 0.90 0.95 1 2 3 4 5 #Hidden Units Accuracy(Bootstrap) Weight Decay q 0e+00 1e?01 1e?04
Controlling Crossvalidation tr <- trainControl(method = ｀cv¨, number = 5) m3 <- train(Species ~ ., iris, trControl = tr, method = ｀nnet¨)
Try more hyperparameter values m4 <- train(Species ~ ., iris, tuneLength = 10, method = ｀nnet¨) q q q q q q q q q q 0.84 0.88 0.92 0.96 5 10 15 #Hidden Units Accuracy(Bootstrap) Weight Decay q 0.0000000000 0.0001000000 0.0002371374 0.0005623413 0.0013335214 0.0031622777 0.0074989421 0.0177827941 0.0421696503 0.1000000000
Use chosen hyperparameter values m5 <- train(Species ~ ., iris, tuneGrid = expand.grid(size=c(1,5,10,20), decay=seq(0,0.1,0.01)), method = ｀nnet¨) q q q q q q q q q q q 0.80 0.85 0.90 0.95 0.000 0.025 0.050 0.075 0.100 Weight Decay Accuracy(Bootstrap) #Hidden Units q 1 5 10 20
Contributions Add your own models. Share by github pull request. But aim is for devs to keep package up to date.
2 0 2 0 S c i e n c e N?OZ
Who develops zoon? Tom August Me Nick Golding Emiel van Loon David Gavaghan Greg McInerny
What does caret do? q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q Sp. Absent Sp. Present 0.00 0.25 0.50 0.75 Feature / predictor variable Class
A basic work?ow work1 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKAir, process = OneHundredBackground, model = RandomForest, output = PrintMap ) q q q 0.935 0.940 0.945 50 75 100 125 150 # Boosting Iterations Accuracy(Bootstrap) Max Tree Depth q 1 2 3
A di?erent work?ow work2 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKBioclim, process = Background(n = 500), model = RandomForest, output = Appify)
A di?erent model work3 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKAir, process = OneHundredBackground, model = MaxEnt, output = PrintMap ) q q q q q q q q q q q 0.80 0.85 0.90 0.95 0.000 0.025 0.050 0.075 0.100 Weight Decay Accuracy(Bootstrap) #Hidden Units q 1 5 10 20
caret in zoon work4 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKAir, process = OneHundredBackground, model = MachineLearn(method = ¨nnet¨, tuneLength = 8), output = PrintMap )
Contributions Add your own methods. Share by web form or github. Not the aim for devs to keep package up to date.
Package ecosystems CRAN Bioconductor zoon caret dismo
Package ecosystems CRAN Bioconductor zoon caret dismo User contribu?on Extendability
Any Questions ? Tim C.D. Lucas

More Related Content

Caret and zoon: machine learning, ecology and domain specific package systems

  • 1. Caret and zoon: machine learning, ecology and domain speci?c package systems Tim C.D. Lucas Malaria Atlas Project, BDI, Oxford @timcdlucas @statsforbios timcdlucas@gmail.com
  • 2. Who am I? Malaria Atlas Project at BDI Malaria, maps, geostatistics
  • 3. Who am I? R packages Zoon INLAutils palettetown - my greatest ever achievement
  • 4. Talk overview caret General package for machine learning. Introduction to the package. A domain speci?c package ecosystem? zoon General package for species distribution modelling. What are SDMs? Package overview. Domain speci?c ecosystems Other examples. Are they a good thing?
  • 6. What is machine learning? q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q Not cat Cat 0.00 0.25 0.50 0.75 Feature / predictor variable Class 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 2 4 6 0.25 0.50 0.75 1.00 Feature / predictor variable Response
  • 9. Hyperparameters Number of PCA coordinates Cut-o?s for variable selection x + x2 + x3 + x4 + ...
  • 10. No free lunch No such thing as a universal, ｀best¨ machine learning model.
  • 13. Training a model m1 <- train(Species ~ ., iris, method = ｀gbm¨) q q q 0.935 0.940 0.945 50 75 100 125 150 # Boosting Iterations Accuracy(Bootstrap) Max Tree Depth q 1 2 3
  • 14. Training a di?erent model m2 <- train(Species ~ ., iris, method = ｀nnet¨) q q q 0.75 0.80 0.85 0.90 0.95 1 2 3 4 5 #Hidden Units Accuracy(Bootstrap) Weight Decay q 0e+00 1e?01 1e?04
  • 15. Controlling Crossvalidation tr <- trainControl(method = ｀cv¨, number = 5) m3 <- train(Species ~ ., iris, trControl = tr, method = ｀nnet¨)
  • 16. Try more hyperparameter values m4 <- train(Species ~ ., iris, tuneLength = 10, method = ｀nnet¨) q q q q q q q q q q 0.84 0.88 0.92 0.96 5 10 15 #Hidden Units Accuracy(Bootstrap) Weight Decay q 0.0000000000 0.0001000000 0.0002371374 0.0005623413 0.0013335214 0.0031622777 0.0074989421 0.0177827941 0.0421696503 0.1000000000
  • 17. Use chosen hyperparameter values m5 <- train(Species ~ ., iris, tuneGrid = expand.grid(size=c(1,5,10,20), decay=seq(0,0.1,0.01)), method = ｀nnet¨) q q q q q q q q q q q 0.80 0.85 0.90 0.95 0.000 0.025 0.050 0.075 0.100 Weight Decay Accuracy(Bootstrap) #Hidden Units q 1 5 10 20
  • 18. Contributions Add your own models. Share by github pull request. But aim is for devs to keep package up to date.
  • 19. 2 0 2 0 S c i e n c e N?OZ
  • 20. Who develops zoon? Tom August Me Nick Golding Emiel van Loon David Gavaghan Greg McInerny
  • 21. What does caret do? q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q Sp. Absent Sp. Present 0.00 0.25 0.50 0.75 Feature / predictor variable Class
  • 22. A basic work?ow work1 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKAir, process = OneHundredBackground, model = RandomForest, output = PrintMap ) q q q 0.935 0.940 0.945 50 75 100 125 150 # Boosting Iterations Accuracy(Bootstrap) Max Tree Depth q 1 2 3
  • 23. A di?erent work?ow work2 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKBioclim, process = Background(n = 500), model = RandomForest, output = Appify)
  • 24. A di?erent model work3 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKAir, process = OneHundredBackground, model = MaxEnt, output = PrintMap ) q q q q q q q q q q q 0.80 0.85 0.90 0.95 0.000 0.025 0.050 0.075 0.100 Weight Decay Accuracy(Bootstrap) #Hidden Units q 1 5 10 20
  • 25. caret in zoon work4 <- workflow( occurrence = UKAnophelesPlumbeus, covariate = UKAir, process = OneHundredBackground, model = MachineLearn(method = ¨nnet¨, tuneLength = 8), output = PrintMap )
  • 26. Contributions Add your own methods. Share by web form or github. Not the aim for devs to keep package up to date.
  • 29. Any Questions ? Tim C.D. Lucas
AboutCopyright
? 2025 際際滷