MachineLearningAlgos

Personal reimplementation of some ML algorithm for learning purposes

Contribution/ Author

This repo is only made by myself (Maxence Giraud), I do not seek additional contributors (although you can open an issue if you find a mistake or have a constructive comment) as it has only learning purposes and no other intentions.

If you want to see the work to come and in progress, you can go to the project tab.

Requirements

• NumPy
• SciPy
• Matplotlib (for plots only)
• Gym (RL environment, not directly used in the algorithms)

The requirements can be installed through this simple command :

pip install -r requirements.txt

Some readme have written LaTeX, you can view them locally with some capable reader (e.g. VSCode) or using extensions (e.g. here for Firefox).

Installation

To install, simply clone the project (don't forget the requirements first) :

git clone https://github.com/MaxenceGiraud/MachineLearningAlgos
cd MachineLearningAlgos/

Usage

• The machine learning algorithms are programmed in a similar fashion as sklearn.
• The deep learning framework is conceived in a Keras-like manner.
• The Reinforcement Learning take as input gym-like environments.

import mla
X,y =  load_your_data()

clf = QDA()
clf.fit(X,y)
clf.score(X,y)

##### Unsupevized learning
clust = mla.DBSCAN(eps=5)
clust.fit_predict(X)

##### Deep Learning
from mla import dl

nn = dl.NeuralNetwork(X.shape[1:],loss=dl.MSE())
nn.add(dl.Dense(10,activation=dl.activation.Relu()))
nn.add(dl.Dense(4,activation=dl.activation.Relu()))
nn.add(dl.Dense(1))

nn.fit(X,y)

##### Multi-armed bandits
from mla import mab 

arms = [mab.arms.Bernoulli(0.8),mab.arms.Exponential(2),mab.arms.Gaussian(2.4),mab.arms.Gaussian(1.5)]
bandit = mab.MAB(arms)

T = 100 # Number of trials
ucb = mab.UCB(bandit.nbArms)
ts = mab.ThompsonSampling(bandit.nbArms)
etc = mab.ETC(bandit.nbArms,T)

mab.bandit_env.RunExpes([ucb,ts,etc],bandit,N_exp=10,timeHorizon=T) # Compare the different algorithms

Other examples on how to use the algorithms can be found in the file example.py

To run unittest :

python -m unittest tests/*/*_test.py

Implemented Algorithms

Some explainations of the algorithms can be found in the readme of their folders, for more thorough reviews check the references.
Algorithms that I plan to implement are written in the projects tab.

Machine Leaning

• KNN (classifier+regressor)
• Logistic Regression/ Perceptron
• Least squares linear regression/classification
• Least squares polynomial Regression/classification
• Ridge
• SVM
• CART Decision Tree
• Gaussian process Regressor
• Naive Bayes (Bernoulli, Gaussian and Multinomial)
• LDA/QDA
• Kernel Ridge
• Kernel KNN

Ensemble methods

• Random Forest
• Bagging
• AdaBoost
• Gradient Boosting
• One vs rest classifier
• One vs one classifier

Clustering / Unsupervized Learning

• K-means
• K-medoids
• DBSCAN
• OPTICS
• Gaussian Mixture with EM algorithm
• Spectral Clustering
• Hierarchical Clustering
• Fuzzy C-means
• Kernel K-medoids

Semi Supzevized Learning

• Label Propagation
• Label Spreading

Deep Leaning / Neural Networks

• Neural Network Base/sequential

Layers

• Dense
• Convolution (1d,2d)
• Max/Avg/Min Pooling (1d,2d)
• Flatten
• Reshape
• Dropout

Wrappers

• Autoencoder
• MLP

Optimizers

• Stochastic gradient descent
• Mini batch Gradient descent
• Epsilon-Delta private SGD
• Adam
• Nadam (Nesterov Adam)
• Adagrad

Loss

All the following loss support weights (if given to fit method, it is done automatically)

• MSE
• MAE
• Binary Cross Entropy
• Neg log likelihood

Activation functions

• Linear
• Relu
• Tanh
• Sigmoid
• Leaky Relu
• Softplus
• Softmax

Multi-armed bandit

Credits to Emilie Kaufmann for the bandit environment.

• UCB
• kl UCB
• Explore Then Commit
• Follow the Leader (FTL)
• Thompson sampling
• Linear UCB
• Linear Thompson sampling

For non stationary bandits, see my other repo here.

Reinforcement Learning

• Value iteration
• Q-learning
• Deep Q-learning
• Advantage Actor Critic (Need PyTorch)

Decomposition / Dimensionality reduction

• PCA (for Probabilistic/Bayesian PCA view other repo here)
• Kernel PCA

Kernels

A KernelFusion is also available that allows to combine kernels w.r.t. differents features of the data

• RBF
• Rational Quadratic
• Exp Sin Squared
• Polynomial
• Laplacian
• Chi Squared
• Sigmoid
• Linear
• Cosine Similarity
• Matérn
• (Normalized) Intersection Kernel (between categorical data)

Metrics

Regression

• MSE
• RMSE
• MAE
• Median Absolute error
• R2 score

Classifcation

• Accuracy score
• Zero one loss

References

The references of the algorithms themselves are for the most part written in the docstring of the corresponding class

[1] M. BISHOP, Christopher. Pattern Recognition and Machine Learning. Springer, 2011.
[2] TIBSHIRANI, Robert, HASTIE, Trevor y FRIEDMAN JEROME, . The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Second Edition. Springer, 2016.
[3] P. MURPHY, Kevin. Machine Learning: A Probabilistic Perspective. The MIT Press, 2012.
[4] Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011.
[5] Courses from Master Data Science held at the University of Lille, France