- Learn
- Machine Learning
- Data: the most important part
ML-M02
Data: the most important part
Data decides how good a model can be: its types, labels, exploration, cleaning, features, splits, leakage and class imbalance.
Most of the work in a real ML project, and most of the ways it fails, are about data rather than models. This module covers everything that happens before you train: knowing what kind of data you have, getting trustworthy labels, looking hard at the data, cleaning it, turning it into features, splitting it honestly, and spotting the leaks and imbalances that make offline scores lie.
The lessons are practical. Each one has a small dataset you can explore or code you can run, and each ends with the mistakes that most often turn a promising model into a production disappointment.
Lessons
- ML-02.01 Data types and datasets Data comes as numbers, categories, ordered categories, text, images, audio, time series and graphs. Beginner 10 min
- ML-02.02 Data collection and labeling Labels are often the most expensive part of ML. Beginner 12 min
- ML-02.03 Exploratory data analysis Exploratory data analysis is looking hard at data before modelling. Beginner 15 min
- ML-02.04 Data cleaning Real data is messy. Beginner 12 min
- ML-02.05 Feature engineering Feature engineering turns raw data into inputs a model can use well. Intermediate 18 min
- ML-02.06 Feature selection More features are not always better: irrelevant ones add noise and cost and can hurt generalization. Intermediate 10 min
- ML-02.07 Train, validation and test splits Data is split into a training set to learn from, a validation set to tune choices and a test set used once at the end to estimate real performance. Beginner 10 min
- ML-02.08 Data leakage Data leakage is when information that would not be available at prediction time sneaks into training, producing great offline scores and poor real performance. Intermediate 12 min
- ML-02.09 Imbalanced data When one class is rare, such as fraud at 0.1 percent, a model that always predicts "not fraud" is 99.9 percent accurate and useless. Intermediate 12 min