2018 February) Coursera: How to Win a Data Science Competition: Learn from Top Kagglers Week 1~ Week 3 Summary

• 0. Machine Learning Algorithms
• Linear Model
• Logistic Regression, Support Vector Machine
• Tools) scikit learn, Vowpal Rabbit
• Tree-based
• Decision Tree, Random Forest, GBDT
• Tools) scikit learn, XGBoost, LightGBM
• KNN-based methods
• Tools) scikit learn
• Neural Networks
• Tools) Tensorflow, Keras, Mxnet, Pytorch, Lasagne

 

• 1. Data Check
• Check List
• Target Metric
• Domain Knowledge
• +How data generated

 

• 2. Data Pre-processing : Based on the data we checked, we have to give a slight difference on data
• Feature checking
• Types : Categorical/Numerical/Text .etc
• Date >using gaps
• Coordinates    >using latitudes, longitudes
• Anonymized    >Try to decode/check correlations
• Scaling    > Standardization
• Missing Values(NaNs) >using Isnull(in XGBoost)/-999,-1 etc
• Outliers    >just keep in mind
• Feature generation
• Label Encoding >Mark numbers rather than symbols for readability
• Frequency Encoding >write the specific 
• Feature Extraction
• Text > Vector
• Bags of Words
• Embeddings(~word2vec)
• Text pre-processing
• consideration : lowercase/lemmatization/stemming/stepwords
• Image > Vector

 

• 3. EDA
• Tool : Visualization
• Using histograms, plots
• Explore feature relations

 

• 4. Building Model + Validation
• Building Model
• Metrics Optimization
• Regression metrics
• MSE
• RMSE
• R^2
• MAE
• MSPE
• MAPE
• RMSLE
• Classification metric
• Accuracy
• Logarithmic Loss
• Validation
• Tools
• Holdout
• LOO
• K-fold
• Splitting Data
• Random, Rowwise : must be independent
• Timewise >Moving Window Validation
• by Id
• Combined
• Problems
• Too limited data
• Too diverse / not inconsistent data

 

 

• Problems

1. Treating values that do not present in train data
1. Inference by the data that have the same feature with the value

 

 

About Metrics Optimization