Final Project/Report

1. Shelter Animal Outcomes Presentation/Code (fin):林子祥, 簡子軒,潘星丞,林唯德

2. House Price (fin): 田兆元, 簡子軒, 曾品華, 高淑婷, 江嘉容
3. Bosch Production Line Performance: 林子軒
4. ML of IMDb Movies evaluation (ppt):  林楷文, 梁智傑, 馮獻慶, 林修禾

XGBoost, GBM and boosted trees

XGBoost

• A gentle introduction to XGBoost by Brownlee contains many informative links.

 

A recent (June 2017) FB post of Yuan-Chun Ivan Chang

今天的討論，我突然發覺有些“只是會用現成工具，卻完全不知其所以然”的人大有人在。甚至只用套件中設定的參數，而完全不懂那些參數的意義。我要如何才能確認這些人是真的知道他們在做什麼？學生的話，我可以直接問，別人的學生我怎麼問呢？問了好像也在問他的老師，不是嗎？

A glance of DL

• [DSC 2016] 系列活動：李宏毅 / 一天搞懂深度學習

• 深入淺出DeepLearning (Vivian Chen)

• Deep Learning Summer School, Montreal 2015

• Deep Learning Book

 

• Lessons from 2 Million Machine Learning Models on Kaggle
• 我玩Kaggle的經驗談

ROC, AUC, pAUC

Criteria

• Training/Testing Error: Convenient but rough
• Confusion matrix/Contingency Table: Better but restrictive to discrete classifiers (or probabilistic/score classifers of given thresholds)
• ROC
  • Interpreting diagnostic tests (ROC curve)
  • Fawcett: ROC101, Intro to ROC curve

“Off-the-shelf” machines

Refer to HTF2009, Section 10.7: “Off-the-shelf” procedures for Data Mining

Also Section 10.8: Spam Data. Pay attention to performance comparisons of these machines.

• Two-sample t-test:  Why?
• MeNemar test

Homework: Check 5/22

• Exercise 10.6
• Construct/Redo Table 10.1 using your data and your favorite machines. Prepare a short presentation (10 min or so) based on your new Table 10.1 and performance evaluation/comparison similar to Section 10.8

BD Prediction: Drive 1

R Notebook: Testdrive 0

Homework/Discussion 1

HTF 2009: Exercise 2.1, 2.8; 3.3(a), 3.30; 4.2, 4.3, 4.9
Cehckpoint 4.27(Thr)

Github Link

A Learning Path of (S)ML via R/Python

VS

• Learning R or Python
• Learning Statistics, Machine Learning, Linear Algebra, etc
• 同時學習 vs. 螺旋學習

適合者： 會些統計（如迴歸等），會些語言（日本語 如 C++）

The Path

Read (foundation: books, papers)

• Textbook: Hastie, Tibshirani and Friedman (2009). The Elements of Statistical Learning: Data Mining, Inference and Prediction. 2nd Edition. (aka. ESLII)  Springer-Verlag.
• An Introduction to Statistical Learning with Applications in R (aka. ISLR), Markham’s summary @ R-bloggers; Intro to linear regression (python).*, Warmenhoven’s repo for ISLR python codes

View (slides, docs)

• Ensemble Learning (Attention to p7-p10,  regression as a learning machine)

Play & Hack

• Playgrounds: Kaggle; CrowdAI, DrivenData, CrowdAnalytix (3 Kaggle alternatives)
• More hacks: 25+ websites to find datasets for data science projects
• Path suggested by Analytic Vidhya

* Diagnosis and remedial measures are needed for sound GLM (regression, ANOVA, ANCOVA) statistical analysis, particularly for modern high dimensional data (small n, large p)

把Python當R用

User Background

• Tasks
  • Statistical data analysis: general linear models (regression, ANOVA, ANCOVA), generalized linear models (logistic regression), PCA, Multidimensional Scaling, simulation
  • numerical analysis (numerical integration, optimization)
  • Statistical machine learning: FDA, Boosting variants, SVM, Random forests
• R core: R (download 0-cloud)
• R proc and packages: e.g. glm, glmnet, ada, GAMboost, e1071, randomforest, rpart.  Search MRO/package
• IDE: R , Rstudio (IDE for R)

Why Python?

• Computational musicology: Miditool box (matlab),   few R packages   —> music21
• Kaggle: Kernels
• Opensource +1: Linux, libraoffice, LaTeX, etc

Python <~ Ruser

• Core: Python 2.7.x or Python 3.5.2
• IDE: Jupyter, Spider
• packages : PyPI
• 把Python當R用

Example: DeepBach

• Project/Problem oriented: DeepBach: a Steerable Model for Bach chorales generation

• Github: Ghadjeres/DeepBach
• Result： musicscore

Lesson: 邊學邊做，邊做邊學

The best of times, the worst of times

• CareerCast Best Jobs 2016,  Kaggle
• 破解人工智慧發展實況的過度炒作    Self-driving Cars is a Dangerous HYPE
• Skynet vs. OpenAI