Multiple-Objective Optimization Auto-Tuning

import paradance as para

Load the offline dataset used for Bayesian automatic tuning.

FILE_PATH is the path where the training data is stored. FILE_NAME is the filename.

loader = para.CSVLoader(
    file_path=FILE_PATH,
    file_name=FILE_NAME,
)

For example, three different targets can be fused like this:

formula="targets[0] - 2 * targets[1] + 100 * max(0, targets[2])".

Select the features to be used for hyperparameter search

selected_columns=['pxtr_0', 'pxtr_1', 'pxtr_2', 'pxtr_3', 'pxtr_4']

We should construct a Calculator, to calculate the composite score based on the selected features.

The calculation method for the composite score can be chosen as either sum or product, corresponding to the additive and multiplicative forms of the formulas.

We can also use the free_style type of equation, which supports any form of eval() expression equation.

cal = para.Calculator(
    df=loader.df,
    selected_columns=selected_columns,
    equation_type="product",
)

Determine the Bayesian search space: whether to search for the power term; whether to search for the first-order term of the features.

power = True
first_order = False

Next, construct the optimization target based on the above settings, and determine the direction of optimization, whether to maximize or minimize the fusion target expression.

The optimization process will be recorded in the LOG_FILE file.

ob = para.MultipleObjective(
    direction="minimize",
    weights_num=len(selected_columns),
    formula=formula,
    power=power,
    first_order=first_order,
    log_file=LOG_FILE,
)

Add the appropriate optimizers to the optimization targets, such as the “woauc”/”auc” optimizer in this case.

target_column is the sub-target that this optimizer is used to optimize, such as “watch_t”/”follow”/”click”.

ob.add_calculator(
    calculator=cal,
    flag='woauc',
    target_column='watch_time'
)

ob.add_calculator(
    calculator=cal,
    flag='auc',
    target_column='follow'
)

ob.add_calculator(
    calculator=cal,
    flag='auc',
    target_column='click'
)

Begin optimization, and save the tuning log. n_trials represents the number of rounds for hyperparameter search.

ob.optimize(n_trials=600)