Skip to content
Generic filters
Exact matches only

5 things you don’t know about PyCaret

Moez Ali
From the author of PyCaret

PyCaret is an open source machine learning library in Python to train and deploy supervised and unsupervised machine learning models in a low-code environment. It is known for its ease of use and efficiency.

In comparison with the other open source machine learning libraries, PyCaret is an alternate low-code library that can be used to replace hundreds of lines of code with a few words only.

If you haven’t used PyCaret before or would like to learn more, a good place to start is here.

In unsupervised machine learning the “n parameter” i.e. the number of clusters for clustering experiments, the fraction of the outliers in anomaly detection, and the number of topics in topic modeling, is of fundamental importance.

When the eventual objective of the experiment is to predict an outcome (classification or regression) using the results from the unsupervised experiments, then the tune_model() function in the pycaret.clustering module, the pycaret.anomaly module, and the pycaret.nlp module comes in very handy.

To understand this, let’s see an example using the “Kiva” dataset.

This is a micro-banking loan dataset where each row represents a borrower with their relevant information. Column ‘en’ captures the loan application text of each borrower, and the column ‘status’ represents whether the borrower defaulted or not (default = 1 or no default = 0).

You can use tune_model function in pycaret.nlp to optimize num_topics parameter based on the target variable of supervised experiment (i.e. predicting the optimum number of topics required to improve the prediction of the final target variable). You can define the model for training using estimator parameter (‘xgboost’ in this case). This function returns a trained topic model and a visual showing supervised metrics at each iteration.

The tune_model function in the pycaret.classification module and the pycaret.regression module employs random grid search over pre-defined grid search for hyper-parameter tuning. Here the default number of iterations is set to 10.

Results from tune_model may not necessarily be an improvement on the results from the base models created using create_model. Since the grid search is random, you can increase the n_iter parameter to improve the performance. See example below:

When you initialize the setup function, you will be asked to confirm data types through a user input. More often when you run the scripts as a part of workflows or execute it as remote kernels (for e.g. Kaggle Notebooks), then in such case, it is required to provide the data types programmatically rather than through the user input box.

See example below using “insurance” dataset.

the silent parameter is set to True to avoid input, categorical_features parameter takes the name of categorical columns as string, and numeric_features parameter takes the name of numeric columns as a string.

On many occasions, you have features in dataset that you do not necessarily want to remove but want to ignore for training a machine learning model. A good example would be a clustering problem where you want to ignore certain features during cluster creation but later you need those columns for analysis of cluster labels. In such cases, you can use the ignore_features parameter within the setup to ignore such features.

In the example below, we will perform a clustering experiment and we want to ignore ‘Country Name’ and ‘Indicator Name’.

In classification problems, the cost of false positives is almost never the same as the cost of false negatives. As such, if you are optimizing a solution for a business problem where Type 1 and Type 2 errors have a different impact, you can optimize your classifier for a probability threshold value to optimize the custom loss function simply by defining the cost of true positives, true negatives, false positives and false negatives separately. By default, all classifiers have a threshold of 0.5.

See example below using “credit” dataset.

You can then pass 0.2 as probability_threshold parameter in predict_model function to use 0.2 as a threshold for classifying positive class. See example below: