Lightwood

Lightwood is the default AI engine used in MindsDB. It deals mainly with classification, regression, and time-series problems in machine learning.

By providing it with the input data and problem definition, Lightwood generates predictions following three core steps that include Data pre-processing and cleaning, Feature engineering, and Model building and training. The input data ranges from numbers, dates, categories, text, quantities, arrays, matrices, up to images, audios, and videos (passed as URLs).

​

How It Works

Here is the algorithm followed by Lightwood starting from the input data setup, through model building and training, up to getting predictions.

The input data is pre-processed and each column is assigned a data type. Next, data is converted into features via encoders that transform data into numerical representation used by the model. Finally, a predictive model takes the encoded feature data and outputs a prediction for the target.

Under the hood, the model splits data into the training, validation, and testing sets, with ratios that are dynamic but usually an 80-10-10 ratio. The split is done by default using random sampling without replacement, stratified on the target column. Doing so, it determines the accuracy of the model by evaluating on the held out test set.

Users can either use Lightwood’s default mixers/models or create their own approaches inherited from the BaseMixer class.

To learn more about Lightwood philosophy, follow this link.

​

Accuracy Metrics

Lightwood provides ways to score the accuracy of the model using one of the accuracy functions.

The accuracy functions include mean_absolute_error, mean_squared_error, precision_score, recall_score, and f1_score.

CREATE MODEL model_name
FROM data_source
  (SELECT * FROM table_name)
PREDICT target_column
USING
  accuracy_functions="['accuracy_function']";

You can define the accuracy function of choice in the USING clause of the CREATE MODEL statement.

​

Tuning the Lightwood ML Engine Features

​

Description

In MindsDB, the underlying AutoML models are based on the Lightwood engine by default. This library generates models automatically based on the data and declarative problem definition. But the default configuration can be overridden using the USING statement that provides an option to configure specific parameters of the training process.

In the upcoming version of MindsDB, it will be possible to choose from more ML frameworks. Please note that the Lightwood engine is used by default.

​

Syntax

Here is the syntax:

CREATE MODEL project_name.model_name
FROM data_source
    (SELECT column_name, ... FROM table_name)
PREDICT target_column
USING parameter_key = 'parameter_value';

​

encoders Key

It grants access to configure how each column is encoded. By default, the AutoML engine tries to get the best match for the data.

...
USING encoders.column_name.module = 'value';

To learn more about encoders and their options, visit the Lightwood documentation page on encoders.

​

model Key

It allows you to specify the type of machine learning algorithm to learn from the encoder data.

...
USING model.args = {"key": value};

Here are the model options:

To learn more about all the model options, visit the Lightwood documentation page on mixers.

​

Other Keys Supported by Lightwood in JsonAI

The most common use cases of configuring predictors use encoders and model keys explained above. To see all the available keys, check out the Lightwood documentation page on JsonAI.

​

Example

Here we use the home_rentals dataset and specify particular encoders for some columns and a LightGBM model.

CREATE MODEL mindsdb.home_rentals_model
FROM example_db
    (SELECT * FROM demo_data.home_rentals)
PREDICT rental_price
USING
    encoders.location.module = 'CategoricalAutoEncoder',
    encoders.rental_price.module = 'NumericEncoder',
    encoders.rental_price.args.positive_domain = 'True',
    model.args = {"submodels": [
                    {"module": "LightGBM",
                     "args": {
                          "stop_after": 12,
                          "fit_on_dev": true
                          }
                    }
                ]};