You have a large dataset of observations with a high number of features.
You need to train a multiclass classification model with hyperparameter tuning in a time- and cost-effective way.
Which of the following decisions helps you to reduce training time and save cost?
Click on the arrows to vote for the correct answer
A. B. C. D.Answer: C.
Option A is incorrect because Grid sampling is used during parameter tuning.
It is applicable for discrete model parameters, to sweep over the entire search space.
It requires a lot of time, so use this method only if your budget allows for the exhaustive search.
Option B is incorrect because the default termination policy is “no forced termination” during hyperparameter tuning, which means that the hyperparameter tuning service will let all training runs complete, i.e.
it doesn't serve saving time and cost.
Option C is CORRECT because by using feature selection you include a process of applying statistical tests to inputs.
The goal is to find the columns which are more predictive of the output.
The Filter Based Feature Selection module provides several feature selection algorithms you can choose from.
Reducing number of features can have remarkable effect on the training time of the model.
Option D is incorrect because overfitting occurs when a model fits the training data too well, and as a result, it can't accurately predict on new data.
It is a kind of “model error”, not related with the training time.
Overfitting is not something you can disable.
Reference:
To reduce training time and save cost when training a multiclass classification model with hyperparameter tuning on a large dataset with a high number of features, the following decision could be made:
Answer: A. Use Grid sampling
Explanation:
Grid sampling is a hyperparameter tuning technique that involves defining a range of values for each hyperparameter of the model, and then training the model with every possible combination of those values in a grid. It is an efficient way to search through the hyperparameter space and find the best combination of hyperparameters for the model.
By using grid sampling, we can explore the hyperparameter space in a systematic way, which helps us to identify the best combination of hyperparameters that give the best performance. This approach saves time and cost by avoiding random search that may take longer to find the best hyperparameters.
On the other hand, default termination policy, filter-based feature selection, and disabling overfitting may not help to reduce training time and save cost as follows:
Default Termination Policy: This policy specifies when to stop the training job based on a metric such as the primary metric, training loss, or validation loss. The default policy stops the training job when the primary metric stops improving for a certain number of epochs. However, this policy does not reduce the training time or save cost since it does not affect the training process itself.
Filter Based Feature Selection: This technique involves selecting a subset of the most relevant features from the dataset based on statistical measures such as correlation or mutual information. Although it can reduce the number of features, it may not always improve the performance of the model, and it does not necessarily reduce the training time or save cost.
Disable overfitting: Overfitting occurs when the model learns the training data too well and becomes less effective in making accurate predictions on new data. Regularization techniques such as L1 and L2 regularization or dropout can be used to prevent overfitting. However, disabling overfitting altogether can result in a model that is not robust enough to generalize to new data, and may not necessarily reduce training time or save cost.
Therefore, the most effective approach to reduce training time and save cost for a multiclass classification model with hyperparameter tuning is to use grid sampling.