Talk:Hyperparameter value pruning in Bayesian Optimization

Overall, this is a wonderful page on using SHAP for hp value pruning in Bayesian Optimization. I learnt how by computing SHAP values we can effectively prune or remove low values from the search space, leading to faster convergence and improved efficiency of the process. I have a few clarification questions.

1. What is the Gini’s importance score?

2. Though the experiments are well explained and seem promising, could you provide more insights on the main motivation and why optimizing the behaviour of the Apache Spark application is important? (This might sound like a naive question but it will help understand the overall motivation better)

Thank you for your valuable feedback.

I added a new subsection and explained how the Ginis score is computed with an example.

I also added more text in the motivation regarding why Spark tuning is necessary and why it is an important problem to explore.

The article was delightful to read and a well-suited topic for our class. It was interesting to learn how SHAP for hyper-parameter value pruning in Bayesian Optimization can help to reduce dimensionality of the hyper-parameter search space. Here are a few suggestions.

• It seems unclear what Gini's importance score is and how is it calculated?
• The experimental section was thoroughly explained, however I missed which datasets were used (article mentioned size 78GB and 321 GB respectively).

Thank you for your valuable feedback.

I added a new subsection and explained how the Ginis score is computed with an example.

Also, I added a citation to how Spark's datasets were executed and more explanation in that regard.