Code for Kaggle Data Science Competitions
Kaggler
Kaggler is a Python package for lightweight online machine learning algorithms and utility functions for ETL and data analysis. It is distributed under the MIT License.
Its online learning algorithms are inspired by Kaggle user tinrtgu’s code. It uses the sparse input format that handles large sparse data efficiently. Core code is optimized for speed by using Cython.
Installation
Dependencies
Python packages required are listed in requirements.txt
- cython
- h5py
- hyperopt
- lightgbm
- ml_metrics
- numpy/scipy
- pandas
- scikit-learn
Using pip
Python package is available at PyPi for pip installation:
pip install -U Kaggler
If installation fails because it cannot find MurmurHash3.h, please add . to
LD_LIBRARY_PATH as described here.
From source code
If you want to install it from source code:
python setup.py build_ext --inplace
python setup.py install
Feature Engineering
One-Hot, Label, Target, Frequency, and Embedding Encoders for Categorical Features
import pandas as pd
from kaggler.preprocessing import OneHotEncoder, LabelEncoder, TargetEncoder, FrequencyEncoder, EmbeddingEncoder
trn = pd.read_csv('train.csv')
target_col = trn.columns[-1]
cat_cols = [col for col in trn.columns if trn[col].dtype == 'object']
ohe = OneHotEncoder(min_obs=100) # grouping all categories with less than 100 occurences
lbe = LabelEncoder(min_obs=100) # grouping all categories with less than 100 occurences
te = TargetEncoder() # replacing each category with the average target value of the category
fe = FrequencyEncoder() # replacing each category with the frequency value of the category
ee = EmbeddingEncoder() # mapping each category to a vector of real numbers
X_ohe = ohe.fit_transform(trn[cat_cols]) # X_ohe is a scipy sparse matrix
trn[cat_cols] = lbe.fit_transform(trn[cat_cols])
trn[cat_cols] = te.fit_transform(trn[cat_cols])
trn[cat_cols] = fe.fit_transform(trn[cat_cols])
X_ee = ee.fit_transform(trn[cat_cols], trn[target_col]) # X_ee is a numpy matrix
tst = pd.read_csv('test.csv')
X_ohe = ohe.transform(tst[cat_cols])
tst[cat_cols] = lbe.transform(tst[cat_cols])
tst[cat_cols] = te.transform(tst[cat_cols])
tst[cat_cols] = fe.transform(tst[cat_cols])
X_ee = ee.transform(tst[cat_cols])
Denoising AutoEncoder (DAE)
For reference for DAE, please check out Vincent et al. (2010), “Stacked Denoising Autoencoders”.
import pandas as pd
from kaggler.preprocessing import DAE
trn = pd.read_csv('train.csv')
tst = pd.read_csv('test.csv')
target_col = trn.columns[-1]
cat_cols = [col for col in trn.columns if trn[col].dtype == 'object']
num_cols = [col for col in trn.columns if col not in cat_cols + [target_col]]
# Default DAE with only the swapping noise and a single encoder/decoder pair.
dae = DAE(cat_cols=cat_cols, num_cols=num_cols, n_encoding=128)
X = dae.fit_transform(pd.concat([trn, tst], axis=0)) # encoding input features into the encoding vectors with size of 128
# Stacked DAE with the Gaussian noise, swapping noise and zero masking in 3 pairs of the encoder/decoder.
sdae = DAE(cat_cols=cat_cols, num_cols=num_cols, n_encoding=128, n_layer=3,
noise_std=.05, swap_prob=.2, mask_prob=.1)
X = sdae.fit_transform(pd.concat([trn, tst], axis=0))
# Supervised DAE with the Gaussian noise, swapping noise and zero masking in 3 encoders in the encoder/decoder pair.
sdae = SDAE(cat_cols=cat_cols, num_cols=num_cols, n_encoding=128, n_encoder=3,
noise_std=.05, swap_prob=.2, mask_prob=.1)
X = sdae.fit_transform(trn, trn[target_col])
AutoML
Feature Selection & Hyperparameter Tuning
import pandas as pd
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from kaggler.metrics import auc
from kaggler.model import AutoLGB
RANDOM_SEED = 42
N_OBS = 10000
N_FEATURE = 100
N_IMP_FEATURE = 20
X, y = make_classification(n_samples=N_OBS,
n_features=N_FEATURE,
n_informative=N_IMP_FEATURE,
random_state=RANDOM_SEED)
X = pd.DataFrame(X, columns=['x{}'.format(i) for i in range(X.shape[1])])
y = pd.Series(y)
X_trn, X_tst, y_trn, y_tst = train_test_split(X, y,
test_size=.2,
random_state=RANDOM_SEED)
model = AutoLGB(objective='binary', metric='auc')
model.tune(X_trn, y_trn)
model.fit(X_trn, y_trn)
p = model.predict(X_tst)
print('AUC: {:.4f}'.format(auc(y_tst, p)))
Ensemble
Netflix Blending
import numpy as np
from kaggler.ensemble import netflix
from kaggler.metrics import rmse
# Load the predictions of input models for ensemble
p1 = np.loadtxt('model1_prediction.txt')
p2 = np.loadtxt('model2_prediction.txt')
p3 = np.loadtxt('model3_prediction.txt')
# Calculate RMSEs of model predictions and all-zero prediction.
# At a competition, RMSEs (or RMLSEs) of submissions can be used.
y = np.loadtxt('target.txt')
e0 = rmse(y, np.zeros_like(y))
e1 = rmse(y, p1)
e2 = rmse(y, p2)
e3 = rmse(y, p3)
p, w = netflix([e1, e2, e3], [p1, p2, p3], e0, l=0.0001) # l is an optional regularization parameter.
Algorithms
Currently algorithms available are as follows:
Online learning algorithms
- Stochastic Gradient Descent (SGD)
- Follow-the-Regularized-Leader (FTRL)
- Factorization Machine (FM)
- Neural Networks (NN) - with a single (NN) or two (NN_H2) ReLU hidden layers
- Decision Tree
Batch learning algorithm
- Neural Networks (NN) - with a single hidden layer and L-BFGS optimization
Examples
from kaggler.online_model import SGD, FTRL, FM, NN
# SGD
clf = SGD(a=.01, # learning rate
l1=1e-6, # L1 regularization parameter
l2=1e-6, # L2 regularization parameter
n=2**20, # number of hashed features
epoch=10, # number of epochs
interaction=True) # use feature interaction or not
# FTRL
clf = FTRL(a=.1, # alpha in the per-coordinate rate
b=1, # beta in the per-coordinate rate
l1=1., # L1 regularization parameter
l2=1., # L2 regularization parameter
n=2**20, # number of hashed features
epoch=1, # number of epochs
interaction=True) # use feature interaction or not
# FM
clf = FM(n=1e5, # number of features
epoch=100, # number of epochs
dim=4, # size of factors for interactions
a=.01) # learning rate
# NN
clf = NN(n=1e5, # number of features
epoch=10, # number of epochs
h=16, # number of hidden units
a=.1, # learning rate
l2=1e-6) # L2 regularization parameter
# online training and prediction directly with a libsvm file
for x, y in clf.read_sparse('train.sparse'):
p = clf.predict_one(x) # predict for an input
clf.update_one(x, p - y) # update the model with the target using error
for x, _ in clf.read_sparse('test.sparse'):
p = clf.predict_one(x)
# online training and prediction with a scipy sparse matrix
from kaggler import load_data
X, y = load_data('train.sps')
clf.fit(X, y)
p = clf.predict(X)
Data I/O
Kaggler supports CSV (.csv), LibSVM (.sps), and HDF5 (.h5) file formats:
# CSV format: target,feature1,feature2,...
1,1,0,0,1,0.5
0,0,1,0,0,5
# LibSVM format: target feature-index1:feature-value1 feature-index2:feature-value2
1 1:1 4:1 5:0.5
0 2:1 5:1
# HDF5
- issparse: binary flag indicating whether it stores sparse data or not.
- target: stores a target variable as a numpy.array
- shape: available only if issparse == 1. shape of scipy.sparse.csr_matrix
- indices: available only if issparse == 1. indices of scipy.sparse.csr_matrix
- indptr: available only if issparse == 1. indptr of scipy.sparse.csr_matrix
- data: dense feature matrix if issparse == 0 else data of scipy.sparse.csr_matrix
from kaggler.data_io import load_data, save_data
X, y = load_data('train.csv') # use the first column as a target variable
X, y = load_data('train.h5') # load the feature matrix and target vector from a HDF5 file.
X, y = load_data('train.sps') # load the feature matrix and target vector from LibSVM file.
save_data(X, y, 'train.csv')
save_data(X, y, 'train.h5')
save_data(X, y, 'train.sps')
Documentation
Package documentation is available at here