stories behind kaggle competitions h2o meetup
TRANSCRIPT
kaggle runs public machine learning competitions
we worked with clients/hosts on various types of problems and data of different sizes
my job as a data scientist at kaggle
“data science is not just kaggle competitions”
whyyyy???
machine learning processes
● Business Problem● Collect Data● Transform Data● Dataset Splitting● Evaluation Metric● Feature Extraction
● Feature Selection● Model Training● Model Ensembling● Methodology Selection● Production System● Ongoing Optimization
not every problem can be turned into a kaggle competition
size matters! where bigger is better (most of the time)
data cleaning/formatting:
● easy to make a quick submission● boosts participation● (too) clean data kills creativity
data privacy/anonymization
metric: how do you measure success?
● Classification - AUC/ Logarithmic Loss/Accuracy
● Regression - RMSE/MAE
● Ranking - MAP/NDCG
● Other / Custom
https://www.kaggle.com/wiki/Metrics
the design of a competition shapes how people are going to solve a problem
Splitting dataset
● training/test● public/private
Time series data
data leakage“Deemed ‘one of the top ten data mining mistakes’, leakage is essentially the introduction of information about the data mining target, which should not be legitimately available to mine from”
“the concept of identifying and harnessing leakage has been openly addressed as one of three key aspects for winning data mining competitions”
“Leakage in Data Mining: formulation, detection, and avoidance” S Kaufman et al
do you have thousands of people reviewing your performance at work 24/7?
I do.
1. people make mistakes. honesty is the best policy.
2. crowdsourcing is powerful. anything that can go wrong will go wrong.
