Trang Le
#math graduate. Postdoc fellow with Jason Moore.
TPOT: A Tree-based Pipeline Optimization Tool
Trang Lê
mathematician
postdoctoral researcher @UPenn
amateur runner
About me
PhD Mathematics, Dec 2017, University of Tulsa
Laureate Institute for Brain Research
- imaging, BrainAGE
- use differential privacy to reduce
overfitting in high-d biological data - pseudopotential for fractional-d
now...
Computational Genetics Lab
UPenn
Clean data
Select features
Preprocess features
Construct features
Select classifier
Optimize parameters
Validate model
Raw data
Automate
Typical pipeline
Automate
Open sourced AutoML tools
- auto-sklearn (python) github.com/automl/auto-sklearn
- Bayesian optimzation over a fixed 3-step ML pipeline
-
auto-Weka (java) github.com/automl/autoweka
- similar to auto-sklearn, built on top of Weka
-
H20.ai (java w/ python, scala, R, web GUI) github.com/h2oai/h2o-3
- basic data prep w/ grid/random search over ML algorithms
-
devol (python) github.com/joeddav/devol
- deep learning architecture search via GP
Randy Olson
- DEAP
- Objective:
- maximize pipeline's CV classification performance
- minimize pipeline’s complexity
- Pareto front with NSGA-II
TPOT
Entire data set
Entire data set
PCA
Polynomial features
Combine features
Select k best features
Logistic regression
Multiple copies of the data set can enter the pipeline for analysis
Pipeline operators modify the features
Modified data set flows through the pipeline operators
Final classification is performed on the final feature set
Genetic programming
GP primitives Dataset selectors, Feature selectors & preprocessors, Supervised classifiers
Population sequences of pipeline operators
Generations
Mutation and crossover
(a) insertion mutation
(b) deletion mutation
(c) swap mutation
(d) substitution mutation
(e) crossover
TPOT configs
- Default TPOT
- TPOT light
- TPOT sparse
- TPOT-MDR (Multi-Directional Reduction)
- Classification
- Regression
TPOT Template
Mutation restriction
Complexity reformulation
- Number of pipeline operators
- Flexibility of each operator
- Runtime
- Number of features used in pipeline
- Number of parameters
- By accessing over-fitting: stability of the covariance of predictors, rank differences of importance metrics
Integration with neural nets
Challenges
- preprocessing
- scalability
- computational expense
@trang1618
Thank you!
Jason Moore
Weixuan Fu
Scalable Automated Machine Learning
By Trang Le
Scalable Automated Machine Learning
Presentation on 2018-12-21 at AI Therapeutics.
