一些筆記, 主要內容比較針對ML

Note for Amazon SageMaker Enablement Workshop

Techinical trainer: Cathy Lai

basis ML flow

基本上這張圖解釋了ML在解決問題的整個flow

Machine Learning Mechanisim

ML type

1. unsupervised clustering (ex: 推薦系統) dimensionalitty reduction

2. supervised classification (ex: 股票買賣) regression (ex: 股票價格)

3. Reinforcement

在本堂課主要是針對supervised 目標是用ML分析潛在可能流失的電信客戶

ML tips

1. data visualzation 在Model training之前, 針對輸入資料做checking

Statistic Analysis：
• Usually use libraries like pandas, NumPy or Matplotlib
• Use Python function like: hist(), describe(), crosstab() & select_dtypes().
• Use Ipython hist library：%matplotlib inline

2. data prepartion

3. Data Cleansing 這部分算是收獲最大的, 以前並沒有做過類似的行為

• Disable unused column before model training
• Use pandas functions corr() and scatter_matrix()
• Confirm the algorithm used after data cleansing
• Important: remove 100% correlation column

Which algorithm?

Classification:

• Linear Learner • XGBoost • Factorization Machines • SVMs (Spark, BYO)

Regression:

• Linear Learner • XGBoost • Factorization Machines • SVMs (Spark, BYO)

Recommendations:

• Factorization Machines • Collaborative Filtering (Spark) • Matrix Factorization (BYO)

Forecasting:

• DeepAR • Linear Learner • XGBoost • Prophet (BYO) • ARIMA (BYO) • EST (BYO)

Clustering:

• K-Means • Hotspot Detection (Kinesis Analytics) • DBScan (BYO) • GMMs (BYO)

Dimensionality Reduction/Anomaly Detection:

• PCA • Random Cut Forest (Kinesis Analytics) • t-SNE (BYO) • Manifold Learning (BYO) • Autoencoders (BYO) • SVMs (Spark, BYO)

Hyperparameter tunning

除非很有經驗, 不然似乎只能用暴力破解