Path3.Mod1.b - Automated Machine Learning - Featurization and Models Flashcards
Differences between Training Data, Validation Data and Test Data
- Training Data: Used to train your models
- Validation Data: Used to tune hyperparameters.
- Test Data: AutoML’s recommended model is tested at the end with this data if provided (preview only)
The problem with Validation Data in AutoML
AutoML uses the same data per iteration of Hyperparameter tuning, creating a model evaluation bias as the model continues to improve and fit to the Validation Data.
Define:
- Feature Engineering
- Featurization wrt Feature Engineering
- Using domain knowledge to create features to help ML learn better.
- The collective techniques of scaling and normalization coupled with Feature Engineering.
Define Customizing Featurization and how to enable it
Additional feature engineering techniques (data manipulation) such as encoding and transforms. These can be enabled through ML Studio and the Python SDK
Define Ensemble Models
The combination of multiple models to improve machine learning results and predictive performance (ensemble learning).
V WA, St heter
The two types of Ensemble Model Combining Methods and which uses a Meta-Model
- Voting: predicts on weighted average of either predicted class probabilities (classification) or regression targets.
- Stacking: combines heterogenous models and trains a meta-model based on the output from the individual models.
LR EN
The two default Ensemble Meta-Models (and for which type of ML task)
- LogisticRegression: for Classfication tasks
- ElasticNet: for Regression and Forcasting tasks
Define ONNX Format and its usage
Open Neural Network Exchange, an open format built for representing ML Models. You can convert your models to ONNX format to run on a variety of platforms
https://onnx.ai/
The ONNX Runtime does not support C#, only Python based models. Ergo you cannot run models in C# applications (T/F).
False. The ONNX Runtime also supports C#, so you can use ONNX formatted models in your C# apps
https://onnx.ai/