機器學習技法 (Machine Learning Techniques)

The course extends the fundamental tools in "Machine Learning Foundations" to powerful and practical models by three directions, which includes embedding numerous features, combining predictive features, and distilling hidden features. [這門課將先前「機器學習基石」課程中所學的基礎工具往三個方向延伸為強大而實用的工具。這三個方向包括嵌入大量的特徵、融合預測性的特徵、與萃取潛藏的特徵。]

• 第一講：Linear Support Vector Machine
• more robust linear classification solvable with quadratic programming
• 第二講：Dual Support Vector Machine
• another QP form of SVM with valuable geometric messages and almost no dependence on the dimension of transformation
• 第三講：Kernel Support Vector Machine
• kernel as a shortcut to (transform + inner product): allowing a spectrum of models ranging from simple linear ones to infinite dimensional ones with margin control
• 第四講：Soft-Margin Support Vector Machine
• a new primal formulation that allows some penalized margin violations, which is equivalent to a dual formulation with upper-bounded variables
• 第五講：Kernel Logistic Regression
• soft-classification by an SVM-like sparse model using two-level learning, or by a "kernelized" logistic regression model using representer theorem
• 第六講：Support Vector Regression
• kernel ridge regression via ridge regression + representer theorem, or support vector regression via regularized tube error + Lagrange dual
• 第七講：Blending and Bagging
• blending known diverse hypotheses uniformly, linearly, or even non-linearly; obtaining diverse hypotheses from bootstrapped data
• 第八講：Adaptive Boosting
• "optimal" re-weighting for diverse hypotheses and adaptive linear aggregation to boost weak algorithms
• 第九講：Decision Tree
• recursive branching (purification) for conditional aggregation of simple hypotheses
• 第十講：Random Forest
• bootstrap aggregation of randomized decision trees with automatic validation
• 第十一講：Gradient Boosted Decision Tree
• aggregating trees from functional + steepest gradient descent subject to any error measure
• 第十二講：Neural Network
• automatic feature extraction from layers of neurons with the back-propagation technique for stochastic gradient descent
• 第十三講：Deep Learning
• an early and simple deep learning model that pre-trains with denoising autoencoder and fine-tunes with back-propagation
• 第十四講：Radial Basis Function Network
• linear aggregation of distance-based similarities to prototypes found by clustering
• 第十五講：Matrix Factorization
• linear models of items on extracted user features (or vice versa) jointly optimized with stochastic gradient descent for recommender systems
• 第十六講：Finale
• summary from the angles of feature exploitation, error optimization, and overfitting elimination towards practical use cases of machine learning