How is AutoML enabling Automatic Model Searching?
Follow Quantifiers for Finance/Analytics Articles
What is AutoML?
Automated machine learning (AutoML) is a method through which we can train high-quality machine learning models even without having expertise in machine learning.
Why do we need AutoML?
- Finding the appropriate machine learning model can be difficult. It can get even more complex when there is a specific business scenario.
- All business enterprises might not be experts in ML and using third-party services or hiring Data engineers can be costly.
How does AutoML address these issues?
- AutoML enables everyone to build machine learning models as it bridges the skill gaps.
- Using AutoML, we can reduce the errors caused due to manual work.
- It enables the automation of several tasks.
What are some popular ways to search in AutoML Frameworks?
We have two variety of frameworks available
- To search for possible models from traditional ML algorithms
- To perform Neural Network Search
How to use AutoML?
The results obtained from AutoML can provide guidance on the models that can be used or the set of features that are important. However, it is not advisable to take the model to production directly.
Parameters to be specified while searching model using AutoML
Maximum run time — This specifies the time taken to experiment
Maximum model — This specifies the maximum number of models that should be built
Stopping metrics and tolerance — Here we specify the tolerance criteria and the metrics like RMSE, ROC
Sorting metrics — These help in determining the method through which we can sort the leaderboard which contains information about the suitable models
Exclude algorithms — We can decide the algorithms that we want to exclude during the model search. For example — we can exclude deep learning models
Include algorithms — We can decide the algorithms we want to include during the model search. For example — we can include Decision tree and XGBoost
Preprocessing — We can mention steps we want to perform in preprocessing data. These could be one hot encoding, scaling etc.
