pspso 0.1.3

pspso is a python package for selecting machine learning algorithms parameters.

Welcome to pspso's documentation!

Overview and Installation

Overview

pspso is a python library for selecting machine learning algorithms parameters. The first version supports two single algorithms: Multi-Layer Perceptron (MLP) and Support Vector Machine (SVM). It supports two ensembles: Extreme Gradient Boosting (XGBoost) and Gradient Boosting Decision Trees (GBDT).

Two types of machine learning tasks are supported by pspso:

• Regression.
• Binary classification.

Three scores are supported in the first version of pspso:

• Regression :

  • Root Mean Square Error (RMSE)

• Binary Classication :

  • Area under the Curve (AUC) of the Receiver Operating Characteristic (ROC)
  • Accuracy

Installation

Use the package manager pip to install pspso.

pip install pspso

Usage

MLP Example (Binary Classification)

pspso is used to select the machine learning algorithms parameters. Below is an example for using the pspso to select the parameters of the MLP. pspso handles the MLP random weights intialization issue that may cause losing the best solution in consecutive iterations.

The following example demonstrates the selection process of the MLP parameters. A variable named params was not given by the user. Hence, the default search space of the MLP is loaded. This search space contains five parameters:

params = {"optimizer": ["RMSprop", "adam", "sgd",'adamax','nadam','adadelta'] ,
      "learning_rate":  [0.1,0.3,2],
      "neurons": [1,40,0],
      "hiddenactivation": ['relu','sigmoid','tanh'],
      "activation":['relu','sigmoid','tanh']} 

The task and the score were defined as binary classification and auc respectively. Then, the PSO was used to select the parameters of the MLP. Results are provided back to the user through the print_results() function.

from sklearn.preprocessing import MinMaxScaler
from pspso import pspso
from sklearn import datasets
from sklearn.model_selection import train_test_split

breastcancer = datasets.load_breast_cancer()
data=breastcancer.data#get the breast cancer dataset input features
target=breastcancer.target# target
X_train, X_test, Y_train, Y_test = train_test_split(data, target,test_size=0.1,random_state=42,stratify=target)
normalize = MinMaxScaler(feature_range=(0,1))#normalize input features 
X_train=normalize.fit_transform(X_train)
X_test=normalize.transform(X_test)
X_train, X_val, Y_train, Y_val = train_test_split(X_train, Y_train,test_size=0.15,random_state=42,stratify=Y_train)
p=pspso(estimator='mlp',task='binary classification', score='auc')
pos,cost,duration,model,optimizer=p.fitpspso(X_train,Y_train,X_val,Y_val)
p.print_results()#print the results
testscore=pspso.predict(p.model,p.estimator,p.task,p.score, X_test, Y_test)
print(1-testscore)

In this example, four parameters were examined: optimizer, learning_rate, hiddenactivation, and activation. The number of neurons in the hidden layer was kept as default.

Output:

Estimator: mlp
Task: binary classification
Selection type: PSO
Number of attempts:50
Total number of combinations: 45360
Parameters:
{'optimizer': 'nadam', 'learning_rate': 0.29, 'neurons': 4, 'hiddenactivation': 'sigmoid', 'activation': 'sigmoid'}
Global best position: [3.8997699  0.28725911 4.21218138 1.41200923 0.84643591]
Global best cost: 0.0
Time taken to find the set of parameters: 160.3374378681183
Number of particles: 5
Number of iterations: 10
0.9867724867724867

XGBoost Example (Binary Classification)

from sklearn.preprocessing import MinMaxScaler
from pspso import pspso
from sklearn import datasets
from sklearn.model_selection import train_test_split

breastcancer = datasets.load_breast_cancer()
data=breastcancer.data#get the breast cancer dataset input features
target=breastcancer.target# target
X_train, X_test, Y_train, Y_test = train_test_split(data, target,test_size=0.1,random_state=42,stratify=target)
normalize = MinMaxScaler(feature_range=(0,1))#normalize input features 
X_train=normalize.fit_transform(X_train)
X_test=normalize.transform(X_test)
X_train, X_val, Y_train, Y_val = train_test_split(X_train, Y_train,test_size=0.15,random_state=42,stratify=Y_train)

params = {
        "learning_rate":  [0.01,0.2,2],
        "max_depth": [1,10,0],
        "n_estimators": [2,200,0],
        "subsample": [0.7,1,1]}
p=pspso(estimator='xgboost',params=params,task='binary classification', score='auc')
pos,cost,duration,model,optimizer=p.fitpspso(X_train,Y_train,X_val,Y_val)
p.print_results()#print the results
testscore=pspso.predict(p.model,p.estimator,p.task,p.score, X_test, Y_test)
print(1-testscore)

XGBoost Example (Regression)

The XGBoost is an implementation of boosting decision trees. Five parameters were utilized for selection: objective, learning rate, maximum depth, number of estimators, and subsample. Three categorical values were selected for the objective parameter. The learning rate parameter values range between 0.01 and 0.2 with 2 decimal point, maximum depth ranges between 1 and 10 with 0 decimal points (1,2,3,4,5,6,7,8,9,10), etc. The task and score are selected as regression and RMSE respectively. The number of particles and number of iterations can be left as default values if needed. Then, a pspso instance is created. By applying the fitpspso function, the selection process is applied. Finally, results are printed back to the user. The best model, best parameters, score, time, and other details will be saved in the created instance for the user to check.

from sklearn.preprocessing import MinMaxScaler
from pspso import pspso
from sklearn import datasets
from sklearn.model_selection import train_test_split

boston_data = datasets.load_boston()
data=boston_data.data
target=boston_data.target

X_train, X_test, Y_train, Y_test = train_test_split(data, target,test_size=0.1,random_state=42)
normalize = MinMaxScaler(feature_range=(0,1))#normalize input features
normalizetarget = MinMaxScaler(feature_range=(0,1))#normalize target

X_train=normalize.fit_transform(X_train)
X_test=normalize.transform(X_test)
Y_train=normalizetarget.fit_transform(Y_train.reshape(-1,1))
Y_test=normalizetarget.transform(Y_test.reshape(-1,1))

X_train, X_val, Y_train, Y_val = train_test_split(X_train, Y_train,test_size=0.25,random_state=42)
params = {
        "objective":['reg:tweedie',"reg:linear","reg:gamma"],
        "learning_rate":  [0.01,0.2,2],
        "max_depth": [1,10,0],
        "n_estimators": [2,200,0],
        "subsample": [0.7,1,1]}
p=pspso(estimator='xgboost',params=params,task='regression', score='rmse')
pos,cost,duration,model,optimizer=p.fitpspso(X_train,Y_train,X_val,Y_val)
p.print_results()#print the results
testscore=pspso.predict(p.model,p.estimator,p.task,p.score, X_test, Y_test)
print(testscore)

User Input

The user is required to select the type of the algorithm ('mlp', 'svm', 'xgboost', 'gbdt'); the task type ('binary classification','regression'), score ('rmse', 'acc', or 'auc'). The user can keep the parameters variable empty, where a default set of parameters and ranges is loaded for each algorithm.

from pspso import pspso
task='binary classification'
score='auc'
p=pspso.pspso('xgboost',None,task,score)

Pspso allows the user to provide a range of parameters for exploration. The parameters vary between each algorithm. Any parameter supported by the Scikit-Learn API for GBDT and XGBoost can be added to the selection process. A set of parameters that contains five XGBoost parameters is shown below. The parameters are encoded in JSON object that consists of key,value pairs:

params = {"objective":['reg:tweedie',"reg:linear","reg:gamma"],
        "learning_rate":  [0.01,0.2,2],
        "max_depth": [1,10,0],
        "n_estimators": [2,200,0],
        "subsample": [0.7,1,1]}

The key can be any parameter belonging to to the algorithm under investigation. The value is a list. Pspso will check the type of the first element in the list, which will determine if the values of the parameter are categorical or numerical.

Categorical Parameters

If the parameter values are categorical, string values are expected to be found in the list, as shown in objective parameter. The values in the list will be automatically mapped into a list of integers, where each integer represents a value in the original list. The order of the values inside the list affect the position of the value in the search space.

Numerical Parameters

If the parameter is numerical, a list of three elements [lb,ub, rv] is expected to be found:

• lb: repesents the lowest value in the search space
• ub: represents the maximum value in the search space
• rv: represents the number of decimal points the parameter values are rounded to before being added for training the algorithm

For e.g if you want pspso to select n_estimators, add the following list [2,200,0]. By that, the lowest n_estimators will be 2, the highest to be examined is 200, and each possible value is rounded to an integer value ( 0 decimal points).

Other parameters

The user is given the chance to handle some of the default parameters such as the number of epochs in the MLP. Although this parameter can be optimized, but its not encouraged. The user can modify this by changing a pspso class instance. For e.g., to change the number of epochs from default to 10 in MLP training:

from pspso import pspso
task='binary classification'
score='auc'
p=pspso.pspso('mlp',None,task,score)# in case of empty set of params (None) default search space is loaded
p.defaultparams['epochs']=10

The verbosity can be modified for any algorithm, which allows showing details of the training process:

from pspso import pspso
task='binary classification'
score='auc'
p=pspso.pspso('mlp',None,task,score)
p.verbosity=1

Early stopping rounds can alos be modified, the user can set a value different to the default value:

from pspso import pspso
task='binary classification'
score='auc'
p=pspso.pspso('xgboost',None,task,score)
p.early_stopping=10

Other parameters such that n_jobs in XGBoost can also be modified before the start of the selection process.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

We are working towards adding the cross validation support that will take the training data and number of folds, then split the records and train each fold. Finally, the average performance is retuned to the user.

We are also working on adding multi-class classification and data oversampling techniques.

Steps for adding another machine learning algorithm

The main reason behind the development of this package is to facilitate the use of the algorithms with a minimum amount of code required, in a way like AutoML. However, the steps to add an algorithm are followed:

• Add a condition in the get_default_search_space() function to include the new algorithm default search space parameters with upper/lower bounds
• Add a default search space based on the algorithm and the task (binary classification or regression) to the get_default_params() function
• Create a function forward_prop_algorithm_name that accepts parameters (similar to forward_prop_gbdt,forward_prop_svm) and returns two variables: the model and fitness value
• Add a condition in the function __f__ to forward the task to the function created in step 3
• Add a condition in the function __predict__ to allow building the model using the function created in step 3

License

Copyright (c) [2020] [Ali Haidar]

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.