Comment on page
- Assess the performance of a linear model on the OTU data in predicting a categorical variable
- Train a model to predict whether an individual has a disease
- Train a model to predict the environment where the sample is taken from
Machine learning models tend to work best with a dataset with a large number of samples
The OTUs will be grouped together (by summing the OTU values) at the selected taxonomic level before the analysis is applied.
The categorical variable that the model should predict.
If there are no values in the dropdown, this likely means that there were no categorical variables found
The type of linear classifier to apply. Choices include:
- Hinge (Linear SVM): Ideal for high-dimensional feature spaces (such as OTU tables). Finds the separating hyperplane that maximizes the margin between targeted classes. Non-linear kernels are not yet available.
- Log (Logistic): Performs similarly to Linear SVMs but may be better when there are fewer samples. A type of generalized linear model that models the parameters as a Bernoulli distribution.
- Huber: Loss function that is less sensitive to outliers
- Squared Hinge: Quadratically penalized hinge loss
- Perceptron: Linear loss by the perceptron algorithm.
Train the model using either by splitting the data into a training and test set or by cross-validating over a specified number of folds.
If set to yes, ensures that the same samples are used as the training set every the model is retrained. This useful to keep the test set untouched.
Define the proportion of the data that should be randomly picked to form a training dataset.
0.5 is recommended.
The maximum number of passes through the training data.
- Assess the predictive performance of your model using the test AUC. Note: Whenever possible, it is still recommended to validate a trained model against an independent dataset (one that is collected outside of your study).
- Tune your model for better performance by looking only at the validation AUC. Tuning refers to changing the configurable parameters to try to achieve a better performance for your dataset. It is important to not tune against the test AUC to ensure you don't overfit your model to the test set.
- The AUC tells you the probability that a randomly sampled positive patient will have a higher predicted score for the positive class than the negative class. The AUC will be shown in a "one-vs-all" format.
- Hover over the training/test error curve generated for the test data
- Save Snapshot: Save the results to the experiment notebook
- Download: Downloads the results as a CSV file
- Share: Creates a shareable link that allows you to share the results with others