AutoMLPipeline

Star History

AutoMLPipeline (AMLP) is a package that makes it trivial to create complex ML pipeline structures using simple expressions. It leverages on the built-in macro programming features of Julia to symbolically process, manipulate pipeline expressions, and makes it easy to discover optimal structures for machine learning regression and classification.

To illustrate, here is a pipeline expression and evaluation of a typical machine learning workflow that extracts numerical features (numf) for ica (Independent Component Analysis) and pca (Principal Component Analysis) transformations, respectively, concatenated with the hot-bit encoding (ohe) of categorical features (catf) of a given data for rf (Random Forest) modeling:

julia model = (catf |> ohe) + (numf |> pca) + (numf |> ica) |> rf fit!(model,Xtrain,Ytrain) prediction = transform!(model,Xtest) score(:accuracy,prediction,Ytest) crossvalidate(model,X,Y,"balanced_accuracy_score") Just take note that + has higher priority than |> so if you are not sure, enclose the operations inside parentheses. ```julia

these two expressions are the same

a |> b + c; a |> (b + c)

these two expressions are the same

a + b |> c; (a + b) |> c ```

Please read this AutoMLPipeline Paper for benchmark comparisons.

• JuliaCon Proceedings:

Recorded Video/Conference Presentations:

• 2024 JuliaCon (End-to-End AI with Julia, K0s, and Argo Workflow)
• 2023 JuliaCon (Wrapping Up Offline RL as Part of AutoMLPipeline Workflow)
• 2022 JuliaCon (Distributed AutoML Pipeline Search in PC/RasPi K8s Cluster)
• 2021 JuliaCon (Finding an Effective Strategy for AutoML Pipeline Optimization)
• 2021 PyData Ireland Meetup (Symbolic ML Pipeline Expression and Benchmarking)
• 2020 JuliaCon (AutoMLPipeline: A ToolBox for Building ML Pipelines)

Related Video/Conference Presentations:

• 2021 JuliaCon (Lale in Julia: A package for semi-automated data science)
• 2019 JuliaCon (TSML: Time Series Machine Learning Pipeline)
• 2021 OpenSource Guild in IBM (Overview of HPC and Data Science in Julia Programming with AutoML)

More examples can be found in the examples folder including optimizing pipelines by multi-threading or distributed computing.

Motivations

The typical workflow in machine learning classification or prediction requires some or combination of the following preprocessing steps together with modeling: - feature extraction (e.g. ica, pca, svd) - feature transformation (e.g. normalization, scaling, ohe) - feature selection (anova, correlation) - modeling (rf, adaboost, xgboost, lm, svm, mlp)

Each step has several choices of functions to use together with their corresponding parameters. Optimizing the performance of the entire pipeline is a combinatorial search of the proper order and combination of preprocessing steps, optimization of their corresponding parameters, together with searching for the optimal model and its hyper-parameters.

Because of close dependencies among various steps, we can consider the entire process to be a pipeline optimization problem (POP). POP requires simultaneous optimization of pipeline structure and parameter adaptation of its elements. As a consequence, having an elegant way to express pipeline structure can help lessen the complexity in the management and analysis of the wide-array of choices of optimization routines.

The target of future work will be the implementations of different pipeline optimization algorithms ranging from evolutionary approaches, integer programming (discrete choices of POP elements), tree/graph search, and hyper-parameter search.

Package Features

• Symbolic pipeline API for easy expression and high-level description of complex pipeline structures and processing workflow
• Common API wrappers for ML libs including Scikitlearn, DecisionTree, etc
• Easily extensible architecture by overloading just two main interfaces: fit! and transform!
• Meta-ensembles that allow composition of ensembles of ensembles (recursively if needed) for robust prediction routines
• Categorical and numerical feature selectors for specialized preprocessing routines based on types

Installation

AutoMLPipeline is in the Julia Official package registry. The latest release can be installed at the Julia prompt using Julia's package management which is triggered by pressing ] at the julia prompt: julia julia> ] pkg> update pkg> add AutoMLPipeline

Sample Usage

Below outlines some typical way to preprocess and model any dataset.

1. Load Data, Extract Input (X) and Target (Y)

```julia

Make sure that the input feature is a dataframe and the target output is a 1-D vector.

using AutoMLPipeline profbdata = getprofb() X = profbdata[:,2:end] Y = profbdata[:,1] |> Vector; head(x)=first(x,5) head(profbdata) ```

julia 5×7 DataFrame. Omitted printing of 1 columns │ Row │ Home.Away │ Favorite_Points │ Underdog_Points │ Pointspread │ Favorite_Name │ Underdog_name │ │ │ String │ Int64 │ Int64 │ Float64 │ String │ String │ ├─────┼───────────┼─────────────────┼─────────────────┼─────────────┼───────────────┼───────────────┤ │ 1 │ away │ 27 │ 24 │ 4.0 │ BUF │ MIA │ │ 2 │ at_home │ 17 │ 14 │ 3.0 │ CHI │ CIN │ │ 3 │ away │ 51 │ 0 │ 2.5 │ CLE │ PIT │ │ 4 │ at_home │ 28 │ 0 │ 5.5 │ NO │ DAL │ │ 5 │ at_home │ 38 │ 7 │ 5.5 │ MIN │ HOU │

2. Load Filters, Transformers, and Learners

```julia using AutoMLPipeline

Decomposition

pca = skoperator("PCA") fa = skoperator("FactorAnalysis") ica = skoperator("FastICA")

Scaler

rb = skoperator("RobustScaler") pt = skoperator("PowerTransformer") norm = skoperator("Normalizer") mx = skoperator("MinMaxScaler") std = skoperator("StandardScaler")

categorical preprocessing

ohe = OneHotEncoder()

Column selector

catf = CatFeatureSelector() numf = NumFeatureSelector() disc = CatNumDiscriminator()

Learners

rf = skoperator("RandomForestClassifier") gb = skoperator("GradientBoostingClassifier") lsvc = skoperator("LinearSVC") svc = skoperator("SVC") mlp = skoperator("MLPClassifier") ada = skoperator("AdaBoostClassifier") sgd = skoperator("SGDClassifier") skrfreg = skoperator("RandomForestRegressor") skgbreg = skoperator("GradientBoostingRegressor") jrf = RandomForest() tree = PrunedTree() vote = VoteEnsemble() stack = StackEnsemble() best = BestLearner() ```

Note: You can get a listing of available Preprocessors and Learners by invoking the function: - skoperator()

3. Filter categories and hot-encode them

julia pohe = catf |> ohe tr = fit_transform!(pohe,X,Y) head(tr)

julia 5×56 DataFrame. Omitted printing of 47 columns │ Row │ x1 │ x2 │ x3 │ x4 │ x5 │ x6 │ x7 │ x8 │ x9 │ │ │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ ├─────┼─────────┼─────────┼─────────┼─────────┼─────────┼─────────┼─────────┼─────────┼─────────┤ │ 1 │ 1.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ │ 2 │ 0.0 │ 1.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ │ 3 │ 0.0 │ 0.0 │ 1.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ │ 4 │ 0.0 │ 0.0 │ 0.0 │ 1.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ │ 5 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │ 1.0 │ 0.0 │ 0.0 │ 0.0 │ 0.0 │

4. Numerical Feature Extraction Example

4.1 Filter numeric features, compute ica and pca features, and combine both features

julia pdec = (numf |> pca) + (numf |> ica) tr = fit_transform!(pdec,X,Y) head(tr)

julia 5×8 DataFrame │ Row │ x1 │ x2 │ x3 │ x4 │ x1_1 │ x2_1 │ x3_1 │ x4_1 │ │ │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ ├─────┼──────────┼──────────┼──────────┼──────────┼────────────┼────────────┼────────────┼────────────┤ │ 1 │ 2.47477 │ 7.87074 │ -1.10495 │ 0.902431 │ 0.0168432 │ 0.00319873 │ -0.0467633 │ 0.026742 │ │ 2 │ -5.47113 │ -3.82946 │ -2.08342 │ 1.00524 │ -0.0327947 │ -0.0217808 │ -0.0451314 │ 0.00702006 │ │ 3 │ 30.4068 │ -10.8073 │ -6.12339 │ 0.883938 │ -0.0734292 │ 0.115776 │ -0.0425357 │ 0.0497831 │ │ 4 │ 8.18372 │ -15.507 │ -1.43203 │ 1.08255 │ -0.0656664 │ 0.0368666 │ -0.0457154 │ -0.0192752 │ │ 5 │ 16.6176 │ -6.68636 │ -1.66597 │ 0.978243 │ -0.0338749 │ 0.0643065 │ -0.0461703 │ 0.00671696 │

4.2 Filter numeric features, transform to robust and power transform scaling, perform ica and pca, respectively, and combine both

julia ppt = (numf |> rb |> ica) + (numf |> pt |> pca) tr = fit_transform!(ppt,X,Y) head(tr)

julia 5×8 DataFrame │ Row │ x1 │ x2 │ x3 │ x4 │ x1_1 │ x2_1 │ x3_1 │ x4_1 │ │ │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ Float64 │ ├─────┼─────────────┼─────────────┼────────────┼───────────┼───────────┼──────────┼────────────┼───────────┤ │ 1 │ -0.00308891 │ -0.0269009 │ -0.0166298 │ 0.0467559 │ -0.64552 │ 1.40289 │ -0.0284468 │ 0.111773 │ │ 2 │ 0.0217799 │ -0.00699717 │ 0.0329868 │ 0.0449952 │ -0.832404 │ 0.475629 │ -1.14881 │ -0.01702 │ │ 3 │ -0.115577 │ -0.0503802 │ 0.0736173 │ 0.0420466 │ 1.54491 │ 1.65258 │ -1.35967 │ -2.57866 │ │ 4 │ -0.0370057 │ 0.0190459 │ 0.065814 │ 0.0454864 │ 1.32065 │ 0.563565 │ -2.05839 │ -0.74898 │ │ 5 │ -0.0643088 │ -0.00711682 │ 0.0340452 │ 0.0459816 │ 1.1223 │ 1.45555 │ -0.88864 │ -0.776195 │

5. A Pipeline for the Voting Ensemble Classification

```julia

take all categorical columns and hot-bit encode each,

concatenate them to the numerical features,

and feed them to the voting ensemble

using AutoMLPipeline.Utils pvote = (catf |> ohe) + (numf) |> vote pred = fittransform!(pvote,X,Y) sc=score(:accuracy,pred,Y) println(sc) crossvalidate(pvote,X,Y,"accuracyscore") ```

julia fold: 1, 0.5373134328358209 fold: 2, 0.7014925373134329 fold: 3, 0.5294117647058824 fold: 4, 0.6716417910447762 fold: 5, 0.6716417910447762 fold: 6, 0.6119402985074627 fold: 7, 0.5074626865671642 fold: 8, 0.6323529411764706 fold: 9, 0.6268656716417911 fold: 10, 0.5671641791044776 errors: 0 (mean = 0.6057287093942055, std = 0.06724940684190235, folds = 10, errors = 0) Note: crossvalidate() supports the following sklearn's performance metric

classification:

• accuracy_score, balanced_accuracy_score, cohen_kappa_score
• jaccard_score, matthews_corrcoef, hamming_loss, zero_one_loss
• f1_score, precision_score, recall_score, #### regression:
• mean_squared_error, mean_squared_log_error
• mean_absolute_error, median_absolute_error
• r2_score, max_error, mean_poisson_deviance
• mean_gamma_deviance, mean_tweedie_deviance,
• explained_variance_score

6. Use @pipelinex instead of @pipeline to print the corresponding function calls in 6

```julia julia> @pipelinex (catf |> ohe) + (numf) |> vote :(Pipeline(ComboPipeline(Pipeline(catf, ohe), numf), vote))

another way is to use @macroexpand with @pipeline

julia> @macroexpand @pipeline (catf |> ohe) + (numf) |> vote :(Pipeline(ComboPipeline(Pipeline(catf, ohe), numf), vote)) ```

7. A Pipeline for the Random Forest (RF) Classification

```julia

compute the pca, ica, fa of the numerical columns,

combine them with the hot-bit encoded categorical features

and feed all to the random forest classifier

prf = (numf |> rb |> pca) + (numf |> rb |> ica) + (numf |> rb |> fa) + (catf |> ohe) |> rf pred = fittransform!(prf,X,Y) score(:accuracy,pred,Y) |> println crossvalidate(prf,X,Y,"accuracyscore") ```

julia fold: 1, 0.6119402985074627 fold: 2, 0.7611940298507462 fold: 3, 0.6764705882352942 fold: 4, 0.6716417910447762 fold: 5, 0.6716417910447762 fold: 6, 0.6567164179104478 fold: 7, 0.6268656716417911 fold: 8, 0.7058823529411765 fold: 9, 0.6417910447761194 fold: 10, 0.6865671641791045 errors: 0 (mean = 0.6710711150131694, std = 0.04231869797446545, folds = 10, errors = 0)

8. A Pipeline for the Linear Support Vector for Classification (LSVC)

julia plsvc = ((numf |> rb |> pca)+(numf |> rb |> fa)+(numf |> rb |> ica)+(catf |> ohe )) |> lsvc pred = fit_transform!(plsvc,X,Y) score(:accuracy,pred,Y) |> println crossvalidate(plsvc,X,Y,"accuracy_score")

julia fold: 1, 0.6567164179104478 fold: 2, 0.7164179104477612 fold: 3, 0.8235294117647058 fold: 4, 0.7164179104477612 fold: 5, 0.7313432835820896 fold: 6, 0.6567164179104478 fold: 7, 0.7164179104477612 fold: 8, 0.7352941176470589 fold: 9, 0.746268656716418 fold: 10, 0.6865671641791045 errors: 0 (mean = 0.7185689201053556, std = 0.04820829087095355, folds = 10, errors = 0)

9. A Pipeline for Random Forest Regression

julia iris = getiris() Xreg = iris[:,1:3] Yreg = iris[:,4] |> Vector pskrfreg = (catf |> ohe) + (numf) |> skrf_reg res=crossvalidate(pskrfreg,Xreg,Yreg,"mean_absolute_error",10)

julia fold: 1, 0.1827433333333334 fold: 2, 0.18350888888888886 fold: 3, 0.11627222222222248 fold: 4, 0.1254152380952376 fold: 5, 0.16502333333333377 fold: 6, 0.10900222222222226 fold: 7, 0.12561111111111076 fold: 8, 0.14243000000000025 fold: 9, 0.12130555555555576 fold: 10, 0.18811111111111098 errors: 0 (mean = 0.1459423015873016, std = 0.030924217263958102, folds = 10, errors = 0)

Note: More examples can be found in the test directory of the package. Since the code is written in Julia, you are highly encouraged to read the source code and feel free to extend or adapt the package to your problem. Please feel free to submit PRs to improve the package features.

10. Performance Comparison of Several Learners

10.1 Sequential Processing

```julia using Random using DataFrames

Random.seed!(1) jrf = RandomForest() tree = PrunedTree() disc = CatNumDiscriminator() ada = skoperator("AdaBoostClassifier") sgd = skoperator("SGDClassifier") std = skoperator("StandardScaler") lsvc = skoperator("LinearSVC")

learners = DataFrame() for learner in [jrf,ada,sgd,tree,lsvc] pcmc = @pipeline disc |> ((catf |> ohe) + (numf |> std)) |> learner println(learner.name[1:end-4]) mean,sd,_ = crossvalidate(pcmc,X,Y,"accuracy_score",10) global learners = vcat(learners,DataFrame(name=learner.name[1:end-4],mean=mean,sd=sd)) end; @show learners; ```

julia learners = 5×3 DataFrame │ Row │ name │ mean │ sd │ │ │ String │ Float64 │ Float64 │ ├─────┼────────────────────────┼──────────┼───────────┤ │ 1 │ rf │ 0.653424 │ 0.0754433 │ │ 2 │ AdaBoostClassifier │ 0.69504 │ 0.0514792 │ │ 3 │ SGDClassifier │ 0.694908 │ 0.0641564 │ │ 4 │ prunetree │ 0.621927 │ 0.0578242 │ │ 5 │ LinearSVC │ 0.726097 │ 0.0498317 │

10.2 Parallel Processing

```julia using Random using DataFrames using Distributed

nprocs() == 1 && addprocs() @everywhere using DataFrames @everywhere using AutoMLPipeline

@everywhere profbdata = getprofb() @everywhere X = profbdata[:,2:end] @everywhere Y = profbdata[:,1] |> Vector;

@everywhere jrf = RandomForest() @everywhere ohe = OneHotEncoder() @everywhere catf = CatFeatureSelector() @everywhere numf = NumFeatureSelector() @everywhere tree = PrunedTree() @everywhere disc = CatNumDiscriminator() @everywhere ada = skoperator("AdaBoostClassifier") @everywhere sgd = skoperator("SGDClassifier") @everywhere std = skoperator("StandardScaler") @everywhere lsvc = skoperator("LinearSVC")

learners = @sync @distributed (vcat) for learner in [jrf,ada,sgd,tree,lsvc] pcmc = disc |> ((catf |> ohe) + (numf |> std)) |> learner println(learner.name[1:end-4]) mean,sd,_ = crossvalidate(pcmc,X,Y,"accuracy_score",10) DataFrame(name=learner.name[1:end-4],mean=mean,sd=sd) end @show learners; ```

```julia From worker 3: AdaBoostClassifier From worker 4: SGDClassifier From worker 5: prunetree From worker 2: rf From worker 6: LinearSVC From worker 4: fold: 1, 0.6716417910447762 From worker 5: fold: 1, 0.6567164179104478 From worker 6: fold: 1, 0.6865671641791045 From worker 2: fold: 1, 0.7164179104477612 From worker 4: fold: 2, 0.7164179104477612 From worker 5: fold: 2, 0.6119402985074627 From worker 6: fold: 2, 0.8059701492537313 From worker 2: fold: 2, 0.6716417910447762 From worker 4: fold: 3, 0.6764705882352942 ....

learners = 5×3 DataFrame │ Row │ name │ mean │ sd │ │ │ String │ Float64 │ Float64 │ ├─────┼────────────────────────┼──────────┼───────────┤ │ 1 │ rf │ 0.647388 │ 0.0764844 │ │ 2 │ AdaBoostClassifier │ 0.712862 │ 0.0471003 │ │ 3 │ SGDClassifier │ 0.710009 │ 0.05173 │ │ 4 │ prunetree │ 0.60428 │ 0.0403121 │ │ 5 │ LinearSVC │ 0.726383 │ 0.0467506 │ ```

11. Automatic Selection of Best Learner

You can use * operation as a selector function which outputs the result of the best learner. If we use the same pre-processing pipeline in 10, we expect that the average performance of best learner which is lsvc will be around 73.0. julia Random.seed!(1) pcmc = disc |> ((catf |> ohe) + (numf |> std)) |> (jrf * ada * sgd * tree * lsvc) crossvalidate(pcmc,X,Y,"accuracy_score",10)

julia fold: 1, 0.7164179104477612 fold: 2, 0.7910447761194029 fold: 3, 0.6911764705882353 fold: 4, 0.7761194029850746 fold: 5, 0.6567164179104478 fold: 6, 0.7014925373134329 fold: 7, 0.6417910447761194 fold: 8, 0.7058823529411765 fold: 9, 0.746268656716418 fold: 10, 0.835820895522388 errors: 0 (mean = 0.7262730465320456, std = 0.060932268798867976, folds = 10, errors = 0)

12. Learners as Transformers

It is also possible to use learners in the middle of expression to serve as transformers and their outputs become inputs to the final learner as illustrated below. julia expr = ( ((numf |> rb)+(catf |> ohe) |> gb) + ((numf |> rb)+(catf |> ohe) |> rf) ) |> ohe |> ada; crossvalidate(expr,X,Y,"accuracy_score")

julia fold: 1, 0.6567164179104478 fold: 2, 0.5522388059701493 fold: 3, 0.7205882352941176 fold: 4, 0.7313432835820896 fold: 5, 0.6567164179104478 fold: 6, 0.6119402985074627 fold: 7, 0.6119402985074627 fold: 8, 0.6470588235294118 fold: 9, 0.6716417910447762 fold: 10, 0.6119402985074627 errors: 0 (mean = 0.6472124670763829, std = 0.053739947087648336, folds = 10, errors = 0) One can even include selector function as part of transformer preprocessing routine: julia pjrf = disc |> ((catf |> ohe) + (numf |> std)) |> ((jrf * ada ) + (sgd * tree * lsvc)) |> ohe |> ada crossvalidate(pjrf,X,Y,"accuracy_score")

julia fold: 1, 0.7164179104477612 fold: 2, 0.7164179104477612 fold: 3, 0.7941176470588235 fold: 4, 0.7761194029850746 fold: 5, 0.6268656716417911 fold: 6, 0.6716417910447762 fold: 7, 0.7611940298507462 fold: 8, 0.7352941176470589 fold: 9, 0.7761194029850746 fold: 10, 0.6865671641791045 errors: 0 (mean = 0.7260755048287972, std = 0.0532393731318768, folds = 10, errors = 0) Note: The ohe is necessary in both examples because the outputs of the learners and selector function are categorical values that need to be hot-bit encoded before feeding to the final ada learner.

13. Tree Visualization of the Pipeline Structure

You can visualize the pipeline by using AbstractTrees Julia package. ```julia

package installation

using Pkg Pkg.update() Pkg.add("AbstractTrees")

load the packages

using AbstractTrees using AutoMLPipeline

expr = @pipelinex (catf |> ohe) + (numf |> pca) + (numf |> ica) |> rf :(Pipeline(ComboPipeline(Pipeline(catf, ohe), Pipeline(numf, pca), Pipeline(numf, ica)), rf))

print_tree(stdout, expr) ```

julia :(Pipeline(ComboPipeline(Pipeline(catf, ohe), Pipeline(numf, pca), Pipeline(numf, ica)), rf)) ├─ :Pipeline ├─ :(ComboPipeline(Pipeline(catf, ohe), Pipeline(numf, pca), Pipeline(numf, ica))) │ ├─ :ComboPipeline │ ├─ :(Pipeline(catf, ohe)) │ │ ├─ :Pipeline │ │ ├─ :catf │ │ └─ :ohe │ ├─ :(Pipeline(numf, pca)) │ │ ├─ :Pipeline │ │ ├─ :numf │ │ └─ :pca │ └─ :(Pipeline(numf, ica)) │ ├─ :Pipeline │ ├─ :numf │ └─ :ica └─ :rf

Extending AutoMLPipeline

If you want to add your own filter or transformer or learner, take note that filters and transformers process the
input features but ignores the output argument. On the other hand, learners process both their input and output arguments during fit! while transform! expects one input argument in all cases. First step is to import the abstract types and define your own mutable structure as subtype of either Learner or Transformer. Next is to import the fit! and transform! functions so that you can overload them. Also, you must load the DataFrames package because it is the main format for data processing. Finally, implement your own fit and transform and export them.

```julia using DataFrames using AutoMLPipeline.AbsTypes

import functions for overloading

import AutoMLPipeline.AbsTypes: fit!, transform!

export the new definitions for dynamic dispatch

export fit!, transform!, MyFilter

define your filter structure

mutable struct MyFilter <: Transformer name::String model::Dict args::Dict function MyFilter(args::Dict()) .... end end

define your fit! function.

function fit!(fl::MyFilter, inputfeatures::DataFrame, target::Vector=Vector()) .... end

define your transform! function

function transform!(fl::MyFilter, inputfeatures::DataFrame)::DataFrame .... end ```

Note that the main format to exchange data is dataframe which requires transform! output to return a dataframe. The features as input for fit! and transform! shall be in dataframe format too. This is necessary so that the pipeline passes the dataframe format consistently to its corresponding filters/transformers/learners. Once you have this transformer, you can use it as part of the pipeline together with the other learners and transformers.

Feature Requests and Contributions

We welcome contributions, feature requests, and suggestions. Here is the link to open an issue for any problems you encounter. If you want to contribute, please follow the guidelines in contributors page.

Help usage

Usage questions can be posted in: - Julia Community - Gitter AutoMLPipeline Community - Julia Discourse forum