Machine Learning
(Note) Contents in bold are included in Coursera Machine Learning lectures. A few topics are not identified---regularized regression, neural networks, and anomaly detection.
- Feature extraction and transformation
- Basic statistics: summary statistics, correlations, hypothesis testing
- Anomaly detection: k-NN (k-Nearest Neighbors)
- Neural networks: perceptron, convolutional neural network
- Optimization: stochastic gradient descent, limited-memory BFGS (L-BFGS, Broyden–Fletcher–Goldfarb–Shanno)
Figure: machine learning algorithm maps: (A) scikit-learn; (B) dlib.
Problems
Learning problems can be roughly categorized as either supervised or unsupervised. Supervised learning builds a statistical model to predict or estimate an output (label) based on some inputs (features): classification if label is categorical, regression if label is quantitative. Unsupervised learning describes the relationships and structure among a set of inputs: dimensionality reduction, clustering.
Other areas of machine learning: Reinforcement learning is concerned with maximizing the reward of a given agent (person, business, etc).
Regression
linear regression
Classification
- Linear classifiers:
- Generative model: linear discriminant analysis (LDA), naive Bayes classifier;
- Discriminative model: Logistic regression (logit), support vector machines (SVM), perceptron;
- Isotonic regression;
Clustering
- k-means clustering;
- hierarchical clustering (dendrogram);
- Gaussian mixture;
- power iteration clustering (PIC);
- latent Dirichlet allocation (LDA);
Standardization is required in case of different units.
Dimensionality Reduction.
- Principal component analysis (PCA): find the (orthogonal) directions in a Euclidean space that successively explain the most sample variance (minimize the residual sum of squares);
- Singular value decomposition (SVD);
Programming Tools
C++/CUDA:
- xgboost: gradient boosting machine (GBM), best GPU performance;
- lightgbm: GBM by Microsoft, best CPU performance;
- Vowpal Wabbit;
- TensorFlow
- Caffe, cuda-convnet2;
JVM (Java, Scala):
- H2O: generalized linear models, gradient boosting machine (also supports random forest), generalized lower rank models, deep neural network;
- Spark: MLlib (not nearly as good);
- Deeplearning4j;
R:
glmnet, randomForest , gbm, e1071 (interface to libsvm), caret,
and more;
Python: scikit-learn sklearn; Pylearn2, Theano;
Benchmark for GLM, RF, GBM: For the algorithms it supports, H2O is the fastest and as accurate on data over 10M records that fit in memory of a single machine. Benchmark for GBM