Learn the distinctions between supervised, unsupervised, semi-supervised, and reinforcement learning based on data and goals.
Supervised learning is broken down into predicting continuous variables with regression and discrete labels with classification.
The principle of 'garbage in, garbage out' highlights the need to analyze and clean data before training any model.
Transform non-numerical string data into a machine-readable format using label, ordinal, and one-hot encoding techniques.
Handle large or small numerical values that can negatively impact training by applying scaling methods like min-max or Z-score normalization.
Address missing values in a dataset by either deleting the affected rows/columns or using imputation to replace them.
A practical example using the adult census dataset demonstrates how to identify and understand biases related to age, sex, and race.
Properly divide your dataset into training and testing sets using strategies like holdout, cross-validation, and stratification to avoid data leakage.
Measure classification model success with accuracy, precision, and F1 score, and regression model success with mean absolute or squared error.
Find the optimal model complexity by balancing the training error and test error to avoid underfitting or overfitting.
Optimize model performance by searching for the best parameters with grid search or randomized search, and explore meta-learning for algorithm selection.
Decision trees offer a transparent, white-box model, but random forests typically provide better performance by combining multiple trees.
A step-by-step Python example shows how to preprocess data, handle missing values, and train decision tree and random forest classifiers using scikit-learn.
Explore the basic building block of neural networks, the perceptron, and see how they are combined into multi-layer perceptron (MLP) architectures.
Go beyond simple MLPs to understand deep neural networks and specialized architectures like CNNs for images and RNNs for language.
Address common issues like overfitting with techniques such as dropout and regularization, and tackle the black box problem with model explainers.
Handle large datasets by using parallelization to train smaller models simultaneously or by leveraging pre-trained components with transfer learning.
Effectively compare models in scikit-learn by using the built-in metrics module and visualizing results with tools like confusion matrices.
Correct for imbalanced target labels in a classification problem by applying oversampling to duplicate instances of the minority class.
Scikit-learn is ideal for beginners due to its ease of use and wide variety of algorithms, though specialized libraries offer deeper customization.
Watch for red flags like poor quality data, excessive missing values, and sensitive information that can compromise model performance and ethics.
29:37Stefan Donsa & Lukas Alber
56:46Jodie Burchell
26:25Lutske De Leeuw
44:11Alexandra Waldherr
52:37Kateryna Hrytsaienko
49:45Kris Howard
32:54Adrian Spataru & Bohdan Andrusyak
43:26Ayon Roy
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
Jungwild GmbH
Understanding Recruitment Group
Barcelona, Spain
LEONHARD WEISS GmbH & Co. KG