Skip to content

Latest commit

 

History

History
236 lines (155 loc) · 6.79 KB

README.rst

File metadata and controls

236 lines (155 loc) · 6.79 KB
Raw

Features Selection

General features selection based on certain machine learning algorithm and evaluation methods

Divesity, Flexible and Easy to use

More features selection method will be included in the future!

Quick Installation

pip3 install MLFeatureSelection

Modulus in version 0.0.7

  • Modulus for selecting features based on greedy algorithm (from MLFeatureSelection import sequence_selection)
  • Modulus for removing features based on features importance (from MLFeatureSelection import importance_selection)
  • Modulus for removing features based on correlation coefficient (from MLFeatureSelection import coherence_selection)
  • Modulus for reading the features combination from log file (from MLFeatureSelection.tools import readlog)

Modulus Usage

  • sequence_selection
from MLFeatureSelection import sequence_selection
from sklearn.linear_model import LogisticRegression

sf = sequence_selection.Select(Sequence = True, Random = True, Cross = True)
sf.ImportDF(df,label = 'Label') #import dataframe and label
sf.ImportLossFunction(lossfunction, direction = 'ascend') #import loss function handle and optimize direction, 'ascend' for AUC, ACC, 'descend' for logloss etc.
sf.InitialNonTrainableFeatures(notusable) #those features that is not trainable in the dataframe, user_id, string, etc
sf.InitialFeatures(initialfeatures) #initial initialfeatures as list
sf.SelectRemoveMode(batch = 2)
sf.GenerateCol() #generate features for selection
sf.clf = LogisticRegression() #set the selected algorithm, can be any algorithm
sf.SetLogFile('record.log') #log file
sf.run(validate) #run with validation function, validate is the function handle of the validation function, return best features combination
  • importance_selection
from MLFeatureSelection import importance_selection
import xgboost as xgb

sf = importance_selection.Select()
sf.ImportDF(df,label = 'Label') #import dataframe and label
sf.ImportLossFunction(lossfunction, direction = 'ascend') #import loss function and optimize direction
sf.InitialFeatures() #initial features, input
sf.SelectRemoveMode(batch = 2)
sf.clf = xgb.XGBClassifier()
sf.SetLogFile('record.log') #log file
sf.run(validate) #run with validation function, return best features combination
  • coherence_selection
from MLFeatureSelection import coherence_selection
import xgboost as xgb

sf = coherence_selection.Select()
sf.ImportDF(df,label = 'Label') #import dataframe and label
sf.ImportLossFunction(lossfunction, direction = 'ascend') #import loss function and optimize direction
sf.InitialFeatures() #initial features, input
sf.SelectRemoveMode(batch = 2)
sf.clf = xgb.XGBClassifier()
sf.SetLogFile('record.log') #log file
sf.run(validate) #run with validation function, return best features combination
  • log reader
from MLFeatureSelection.tools import readlog

logfile = 'record.log'
logscore = 0.5 #any score in the logfile
features_combination = readlog(logfile, logscore)

Function Parameters

sf.ImportDF(df,label)

df: pd.DataFrame, include all features
label: str, name of the label column

sf.ImportLossFunction(lossfunction,direction)

lossfunction: handle of the loss function, function return score as scalar value (logloss, AUC, etc)
direction: 'ascend'/'descend', direction to improve

sf.InitialFeatures(features)

features: list of initial features combination,
          empty list will drive code to start from nothing
          list with all trainable features will drive code
          to start backward searching at the beginning

sf.InitialNonTrainableFeatures(features) #only for sequence selection

features: list of features that not trainable (string, datetime, etc)

sf.GenerateCol(key=None,selectstep=1) #only for sequence selection

key: str for the selected features, only the features with keyword will be seleted,
     default to be None
selectstep: int, value for features selection step, default to be 1

sf.SelectRemoveMode(frac=1,batch=1,key='')

frac: float, percentage of delete features from all features
      default to be 1 as using the batch
batch: int, delete features quantity every iteration
key: str, only delete the features with keyword

sf.SetTimeLimit(TimeLimit)

TimeLimit: float, maximum running time, unit in minute

sf.SetFeaturesLimit(FeaturesLimit)

FeaturesLimit: int, maximum feature quantity

sf.SetClassifier(clf)

clf: classfier or estimator, sklearn, xgboost, lightgbm, etc

sf.SetLogFile(logfile)

logfile: str, log file name

sf.run(validate)

validate: function handle with score and classifier return

def validate(X, y, features, clf, lossfunction):
    """define your own validation function with 5 parameters
    input as X, y, features, clf, lossfunction
    clf is set by SetClassifier()
    lossfunction is import earlier
    features will be generate automatically
    function return score and trained classfier
    """
    clf.fit(X[features],y)
    y_pred = clf.predict(X[features])
    score = lossfuntion(y_pred,y)
    return score, clf

def lossfunction(y_pred, y_test):
    """define your own loss function with y_pred and y_test
    return score
    """
    return np.mean(y_pred == y_test)

DEMO

More examples are added in example folder include:

  • Simple Titanic with 5-fold validation and evaluated by accuracy (demo)
  • Demo for S1, S2 score improvement in JData 2018 predict purchase time competition (demo)
  • Demo for IJCAI 2018 CTR prediction (demo)

PLAN

  • better API introduction will be completed next before the end of 06/2018

This features selection method achieved

  • 1st in Rong360

-- https://github.com/duxuhao/rong360-season2

  • Temporary Top 10 in JData-2018 (Peter Du)
  • 12nd in IJCAI-2018 1st round

Algorithm details (selecting features based on greedy algorithm)

Procedure

Procedure