What kind of models and training methods are used by tangram?

[m-kru]

I am trying to get to know what kind of models and training methods are used by tangram. However, I struggle to find any information. The help message from tangram train --help says nothing on that. I can only see that tangram is training 8 models. In the About info page I can read "... of the gradient boosted decision tree algorithm.", but this information is also very rudimentary.

[isabella]

Hi @m-kru. We should add the information about which models are trained in tangram train --help!

The current default hyperparameter grid is as follows:

Linear Models

learning_rate: [0.1, 0.01] l2_regularization: [1.0, 0.1] max_epochs: 1000

Gradient Boosted Decision Trees:

learning_rate: [0.1, 0.01] l2_regularization: [1.0, 0.1] max_leaf_nodes: 512 max_depth: 50 max_rounds: 1000

Does that answer your question?

[m-kru]

@isabella thanks for your quick reply. What Linear Model is it? Do you mean Linear Regression?

[isabella]

Linear regression trained via Stochastic Gradient Descent. The default batch size is 128.

[lnicola]

@isabella are classification (i.e. non-regression) trees supported? I tried to train a model, but the statistics include the RMSE, which doesn't really make sense for classification.

[isabella]

hi @lnicola. Yes classification models are definitely supported. The most likely explanation you are getting a regreession model is that your target column is numeric. If you would like Tangram to train a classification model given a numeric target column, you can pass a config file to training. e.g. tangram train --config config.json ...

{
    "dataset": {
        "columns": [
            {
                "name": <your target column>,
                "type": "enum",
                "variants": [
                    <variant 1>,
                    <variant 2>,
                    <variant 3>
                                        ...
                ]
            }
                ]
        }
}